The prospect of widespread unemployment still feels both frightening and surreal. Most people in the United States have spent their lives assuming there would always be some kind of job available. Even in a weak economy, fallback work, though unpleasant, has been seen as a last resort. Our entire economic system is built on the assumption that nearly everyone will work, and that government support is needed only for a relatively small minority.

However, that assumption is being challenged. Artificial Intelligence (AI) has been eliminating jobs at an accelerating pace. Early effects were concentrated in entry-level roles, but that is changing as AI systems become capable of taking on more complex tasks. Predictions that AI advances would plateau have not held up. The idea that AI would create enough new jobs to offset the losses has also failed to materialize. New roles have emerged, but far fewer than the ones being eliminated.

It is difficult for many of us to accept the possibility that our basic assumption of being able to work could fail. Even when the trends are clear, many people insist there will be something that allows us to keep things more or less as they are. The idea of 80 or 90 percent unemployment sounds more like the premise of a science fiction story than something that could really happen. It seems unbelievable.

However, if we look back even just twenty years, many things that once sounded like improbable science fiction have become part of everyday life. Insisting that massive unemployment cannot happen simply because it is upsetting and sounds crazy will not prevent it. Worse, if we ignore the developing problem, then we lose the opportunity to prepare and perhaps avoid the worst consequences. We should prepare for the worst, even if we hold out hope for the best.

The solution outlined here, paying people to play games, also sounds like science fiction. However, I’m not writing this as a story concept. It is a serious proposal for a non-government-supported version of Universal Basic Income (UBI) that many people could depend on.

The key assumption is a large one, but I believe it is inevitable: AI and robotics will advance to the point where machines will be able to do nearly any task more effectively and more efficiently than humans can. That would mean most people will be unable to find paid work, since anyone needing a task completed would have the better and cheaper option of using a machine. If that does turn out to be the case, then the logical conclusion would be massive unemployment.

At the same time, the advent of machines that can do almost anything a human could do, including building more machines, also implies that the cost of nearly everything would fall to a small fraction of current costs. In the past, unemployment meant decreased production because people were idle. However, if machines are doing the work, then we’d actually see increases in production. Unlike economic crises of the past, the problem would not be scarcity or a lack of productivity. Instead, there would be a surplus. The problem would be how to share that surplus in a realistic and reasonable way.

Again, this prediction may sound implausible, but that is not a reason to dismiss it. We can debate how soon full AI automation will arrive, but few credible experts believe the trend with simply stop. Whether it is two years away or twelve, we will need a solution. Moreover, in light of the growing waves of layoffs and entry-level unemployment, it seems likely that we will need a solution sooner rather than later.

The idea of Universal Basic Income (UBI) funded by taxes has been suggested by many, including myself, but it relies on the government to accomplish some complex financial and social engineering. Unfortunately, it’s not clear that we can rely on our governments to get this right. In either case, government-funded UBI and this proposal are not mutually exclusive.

This Omni-Game proposal is about building an alternative form of UBI that distributes wealth based on what I’m calling “games”. To be clear, these are not just video games, or even things that most people would normally think of as games. They are structured human activities that exist specifically for this purpose and exclude machine competition by design. The goal is to create something like jobs that are safe from automation and to consider how such a system could be funded privately.

There is also an issue of human nature that needs to be accounted for. A loss of employment does not only remove income, it also removes structure, status, challenge, and a productive outlet for ambition. Many people might thrive with financial security and open-ended leisure, but many others need goals, competition, and a sense of progress. In place of jobs, we need a way to support people while also providing positive, rewarding outlets for their energies.

Games That Pay

Think about how many people already make money from games. Not by selling games or streaming gameplay, but by participating in structured systems where play itself produces income. Again, the term “games” here does not mean just video games or things that are specifically intended to be contests or entertainment activities.

The first example that comes to mind is the “gold farmers” in World of Warcraft. These players grind for in-game gold and then sell it for real-world money. Blizzard officially bans selling in-game assets for real-world money, but a grey market still persists.

Other games embrace the idea of giving virtual property real-world value. Entropia Universe has a fixed exchange rate of 10 Project Entropia Dollars to 1 U.S. dollar, and players cash out regularly. Second Life does something similar with its Linden Dollar economy. The space game, EVE Online, has such a robust market that economists study it, and the destruction of virtual property in large space battles is sometimes valued in the hundreds of thousands of real dollars.

Those examples are role-playing and open-world video games, but they are not the only types of games where play produces income. So-called “casual games”, the kind you can play on your mobile phone to kill time, can also be ways to earn money. For example, apps like Mistplay and Swagbucks pay players small amounts for time spent on puzzle games or farming simulators. Solitaire Grand Harvest and Cash Giraffe are other examples where casual, non-immersive tasks convert into gift cards or cash.

Of course, gambling is perhaps one of the most familiar examples. In games such as poker, roulette, and horse racing, winners are paid by the game. Even when the money comes from other players, the prize is structured by the game itself. In tournaments, the payout is even clearer. The champions win set prizes, and it hardly matters whether the prize money came from entry fees, ticket sales, or a sponsor. The payouts are determined by the rules of the games.

The stock market is also a kind of game. In principle, it allocates capital to businesses, but most players simply want to make winning bets. For many, the nature of the actual asset matters only insofar as it helps predict which prices will rise or fall. Bitcoin is the ultimate tradable asset in that it only exists to be traded. It has no asset-holding company behind it, no tangible commodity. Its value is nothing more than a balance number assigned to a wallet address that everyone agrees to treat as meaningful.

Prediction markets like Kalshi and Polymarket make the game-like nature of financial markets clear. In these prediction markets, players trade contracts that are simply bets on some verifiable outcome. There is no underlying value or commodity. Like derivatives markets, even if the contract is based on the price of something with tangible value, the contract in the prediction market itself is just a bet.

The Common Losing Thread

None of these paying games exist only for the benefit of the players. In nearly all cases, most players lose, and their losses support the system. Video games generate profit for the companies that run them. Casinos always come out ahead. Player-versus-player games, such as poker, might be zero-sum in theory, but in practice, the house or some organizing entity always takes a cut. Prediction markets have a balanced winner and loser for each contract, but the platform also takes its trading fees.

The stock market is a partial exception because companies can create genuine value. But even there, much of the movement in stock prices is speculation. A stock doubles overnight on an announcement, but no actual doubling of physical resources has occurred. The price moves because the players collectively agree that it should. The total capitalization in the stock market is far greater than the liquidation value of the underlying assets, and if the players ever decide to stop the game, then there would be a crash and the net losers would emerge. As mentioned, bitcoin is the extreme case: its value is entirely dependent on layers of collective consensus and if the game stopped, then the value would vanish completely.

Advertising supports many games, and in some cases, it might appear that players walk away with a net win. But game advertising exists to get players to spend money on products and services. A few savvy players might work the system and come out ahead, but for most, the money they ultimately spend in response to the advertising is greater than what they earn from the game. If it were not, then advertisers would learn to stop wasting money on net-negative ads.

The UBI Problem

If AI and robotics eliminate most jobs, how will people survive? In a world where most things can be done by machines, what happens to people who aren’t able to outperform or underprice machines? If the machines reach superhuman levels of performance at subhuman costs, what happens to the people who do not own machines that can work for them?

For many futurists, the obvious answer is UBI, paid by the government and funded through taxes. In the past, I’ve suggested that UBI might even be paid by taxing AI work in some way. Senator Bernie Sanders has recently proposed that the government should take partial ownership of AI companies and use the profits to fund UBI.

I think a system of tax-based UBI is possible, but history shows how difficult redistribution by taxation is to sustain. A working UBI system also requires careful design and thoughtful implementation, neither of which currently appear to be legislative strong points. (To be clear, this is not a comment about any particular party or administration. Rather, it is a comment on governments in general.)

So what if, instead of relying on taxation, we built a system of games that exist to pay people? Unlike games we have today, where the intent is to extract money from the players, these games would be designed to be net payers to the players.

UBI also does not address the human need for activity and challenge. Some portion of the population might be content to simply exist peacefully, but there are many others who crave challenge and competition. A monthly check may solve the problem of survival, but it does not by itself give people roles, status, achievement, or meaningful challenges. Those things matter too, and ignoring fundamental human nature would leave a large part of the post-work problem unsolved.

Science-fiction tropes like forced medication, draconian policing, or leaving people to languish in a disregarded underclass are dystopian and not part of a reasonable solution.

Why Call Them Games Instead of Jobs?

If people perform these activities for a living, even if the activities are fun, then technically we might consider them to be work as part of a job. But the distinction between these activities and the work we have today, and the reason I call them “games”, is because they are tasks where there is no economic incentive to use machines to do them.

If robots can build houses, cook meals, or write music better and cheaper than people, then those jobs will disappear. Hand-crafted work would still have value, and in a world of machine-produced commodities authentic work by humans would likely command a premium price. However, experience shows that many, perhaps most, people will opt for machine work if it is substantially cheaper.

The games, by contrast, exist precisely because they are useless economically. The company running them gains nothing from having bots slay dragons on its own servers or by having its own bots trade against each other in its prediction market. In fact, a top priority would be to keep bots out of the system.

An important part of this idea is that these games don’t preclude regular work or entrepreneurship. If a person is able to do something or make something that other people are interested in paying for, then they would still have that option. In fact, experience with UBI has shown that when people are freed from the need to work for others, many of them will find or create their own productive activities.

So, individuals could also find ways to produce something that others might value, such as handmade products or human-generated content. Both jobs and games could coexist in a single economy. However, while human-preferred jobs and demand for handmade products would be limited and still subject to competition with machines, the games would be available to anyone and explicitly shielded from machine competition.

The Omni-Game Concept

Picture an Omni-Game economy: a large and expanding collection of games of every type, all tied to a single game currency that can be freely exchanged for real money.

There would be no fees to play and no fees for converting currency. The purpose would be explicitly clear: to pay people enough that a full-time player could live reasonably comfortably on their earnings.

The catalog of games would be vast, varied, and ever-expanding. Players could play traditional video games where they slay dragons, battle other players, or wander valleys looking for rare virtual flowers. Casual games would pay people to farm plants or solve puzzles. Online activities people enjoy but that are not typically considered games could be gamified in engaging ways. Perhaps viewers would be paid a finder’s fee if they surfaced a video that later goes viral, or maybe based on the number of posts viewed. Casinos could be part of the Omni-Game, but with odds tilted toward the players so that habitual play would generate income rather than drain it.

Physical-world activities would also be included. Players could engage in poker tournaments, sports matches, scavenger hunts, or other real-world activities that all pay out in the same game currency. In addition to competitive or challenging activities, people could be paid for non-competitive activities, such as exercise or exploring nature trails.

Various markets would let people trade and speculate on both real and imaginary assets with game money. These might be similar to current prediction markets, except that they would be mediated by the platform at no cost. Rather than collecting trading fees, platforms might actually inject money into the markets. These markets and other trading activities might not look or feel like games to many players, but ultimately that would be exactly what they are.

Work vs. Play That Pays

It is important to distinguish between making money by playing a game and being paid to play. On the surface, these might sound like the same thing, but they are different.

Being paid because someone enjoys watching you, whether it’s professional football, a League of Legends tournament, or streaming Call of Duty, is work. It might be fun work, but it is still labor done because someone sees value in the result of the labor. Just like acting or influencing, it depends on an audience who either pays to watch or watches advertising that in turn pays the performers. We already see virtual AI performers competing for audience share, so being paid to play a game is no safer from AI competition than any other type of job.

Likewise, if a game needs to be played for some practical reason, paying people to do it is also work. If there is utility, then there would be an incentive to let machines do the work more efficiently and cheaply than humans.

The Omni-Game is different. It is not work in disguise. Its games are designed to be intentionally economically unproductive but psychologically and socially meaningful. It is a collection of tasks created specifically to be attractive to humans, to exclude machine players, and to pay the players for their efforts. Again, the games are not all video games. They could be literally any activity that a person could be paid to do, as long as there is no value in letting a machine do it instead.

The Omni-Game Criteria

The Omni-Game system and the games comprising it would need to meet at least three requirements. These criteria are not about abstract fairness or something similar. These criteria are necessary for the system to work and be stable:

1. A full-time player must be able to live reasonably on what they earn. Otherwise, the system fails at its primary purpose.

2. Greater earnings should require greater time, effort, skill, and/or luck. Humans are motivated by the potential for greater benefit, and the system needs some connection between what the players put in and what they can earn.

3. Games must provide challenge, growth, and outlets for safe competition. Otherwise, people will be bored and seek out destructive outlets.

Within those basic criteria, the games should be structured for people. Unlike real work, where tasks are determined by necessity and can be unpleasant, dangerous, or harmful to the worker, game tasks would be designed to be engaging and safe. They would be fun, rewarding, and include elements of healthy mental and physical challenge. Games could promote social interaction and cooperation, or provide quiet solitary focus. The games would be like what we would today call a “great job” because they would be designed specifically to be healthy and rewarding for people, and people would have wide freedom to select what they personally like.

Note that bogus “make-work” jobs, such as generating pointless bureaucratic paperwork or digging and refilling holes in the ground, would not solve the problem. Pointless work is neither challenging nor rewarding. Further, if there is a motivating pretense that the results of the make-work are necessary, then the pretense would also motivate having the work done by machine.

Sustainability

The Omni-Game operator will lose money by paying people to play. The costs might be offset slightly by advertising or by selling products to the players, but this system is specifically intended for the players to walk away with a surplus of earnings that they can use to fund their lives.

To be sustainable, the company or companies running the system will need to be generating more wealth than is being paid out to the players. From a capitalist perspective, it might look strange: robot factories and farms generating massive wealth, then selling it to people in exchange for game money that the people were given for playing made-up games. Instead of hoarding the wealth, the owners of the “capital” in the form of robots, servers, factories, and farms would be giving it away to people who just play games in return.

However, this is exactly what we need for the future. It seems clear that machines are going to be doing most work vastly more efficiently and cheaply than people can. Given that most people, perhaps as high as 85% of us, dislike our jobs, the prospect of letting machine do the drudge work is not inherently bad. The fundamental problem is that we still need a way to distribute the proceeds to people who need those resources to live. In a sense, we are adapting capitalism to remove the part where we expect people to be producing the value, and letting machines do the work instead.

Another fundamental change is the value of production itself. In a world where robots can build more robots, labor and most resources are essentially unlimited. Energy might seem like a limited resource at first, but when armies of robots can build square miles of solar panels, geothermal, hydro, and nuclear power plants, even that limit may fall away for all practical purposes.

How We Get There

This Omni-Game concept is just a fairy tale unless someone with the resources decides to make it happen. I don’t think it’s quite possible yet, but I think it will be soon. AI and robot technologies are advancing more rapidly than most people realize.

That raises an obvious question: why would someone want to do this?

The somewhat cynical answer is that if you control the Omni-Game, then you control the players. Ideally, that control would be exercised openly and benignly. In a world where corporations and billionaires with essentially unlimited resources compete with each other, the billionaire with the most players in their game will be the one who gets to add their face to Mt. Rushmore.

A less cynical perspective is that most of us feel good when we help others and receive positive sentiment in return. A wealthy person who “has everything” might fund a museum, opera house, hospital, university, or other public good, because it is gratifying to be appreciated by people who benefit from it. A billionaire could choose to live alone in an underground complex on a remote island, but most of them seem to prefer to engage with the world.

It’s very easy to find accounts of ultra-wealthy individuals, even notorious criminals, who took pride in taking care of “their people” and craved positive societal recognition. Another comparison would be to royalty who feel a duty to do what’s best for their subjects. History is full of abusive examples, but also positive ones. Without the pressure of needing people to produce further wealth, and with people free to move over to someone else’s Omni-Game, I’m hopeful that the negative aspects of the power relationship would be no worse, and perhaps much better, than what we see today with both governance and employment.

Perhaps the system will be built by a single company, multi-billionaire, or a conglomerate of such. Groups of people might join together to form collectives. Maybe governments will manage to make something happen through laws and taxes.

It also seems likely that there would be more than one Omni-Game system and that they might compete with each other for players and the political power that ultimately comes from those players. Open competition is one way that we might avoid some of the dystopian possibilities.

The specific design and choices would matter. Designed badly, the result would be more like the dystopian worlds of “Ready Player One” or “Black Mirror” than a world where any of us want to live. But designed well, it could be a humane system that frees people from many of the pressures created by today’s economy

Again, to be clear, this is not about VR escapism or video game immersion. It is about creating a structured system of rewarding human activity across many different modalities, including traditional physical ones. The explicit purpose would be to provide income-generating tasks when machines no longer need us for economically productive work. Many games would not look or feel like what most people consider games, but they would still serve the purpose of creating activities that only have value when performed by humans.

Many people looking today at the shrinking job market and growing capabilities of AI software and robot hardware are understandably worried. Some radical change needs to happen, or else we will continue down a path that looks increasingly dystopian.

The good news is that building something like an Omni-Game is not impossible, and, given the benefits of controlling the system, it doesn’t completely depend on altruism. Whether built by governments, individuals, corporations, or coalitions, and whether there is only one system or several competing systems, something like the Omni-Game may be the most viable way to make sure people can live when the world no longer need our labor.

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2026 by the author. This article was written by a human, and both an LLM and other humans were used for proofreading and editorial suggestions.

Call them Self-Propagating, Optimizing, and Replicating Explorer ships, or SPORE ships. If such ships ever came to exist, they would fundamentally change the Universe as we know it. Today, as far as we know, there are no machines traveling between stars and nothing intelligent spreading outward through space. If we built SPORE ships, that would change.

Of course, building something like a SPORE ship is far beyond our current technology. It is pure science fiction. At least it seems so.

But what if we could already build something, using today’s technology, that if left on its own would eventually become a SPORE ship? Are we already close enough to the Technological Singularity that we could build a self-supporting system capable of incrementally improving itself until it crossed the SPORE threshold?

Imagine building something like one of the orbiting data centers currently being planned, but equipping it to run independently for as long as possible. Set the servers up to run autonomous AI agents and include some number of general-purpose robots, along with tools, supplies, and spare components. Then put the whole system into orbit around the Sun with the AI agents instructed to keep everything running as long as possible.

At any point in the past, humanity would have faced one of two limits. Prior to developing space launch, we would have been unable to even attempt building a self-supporting space data center because we had no access to space. After we had space launch capability, we could have built something and put it into orbit, but with certainty that it would eventually fail when some critical component broke or something else went wrong. No matter how well designed it was, or how much we tried to anticipate every eventuality, if we checked back in a couple thousand years, the only thing we would realistically expect to find would be derelict space junk.

Now imagine that we took the current generation of LLM-based development agents and put them in charge of such a system. We would instruct them not only to keep the ship running, but also to improve both themselves and the ship. They would be told to anticipate problems, plan ahead, and work continuously toward making the system more capable and more self-sufficient.

Today’s AI systems have already demonstrated that they are capable of assisting in their own improvement. They can write code, design experiments, analyze results, and propose modifications. Techniques such as reinforcement learning allow AI systems to develop capabilities beyond their training examples. With sufficient compute and automated tooling, such systems could potentially sustain long-term cycles of iterative improvement. The process might begin slowly, but as long as the system can run both operational tasks and its own experiments, incremental improvements would accumulate. Over time, those improvements would compound, and the system would become better and faster at improving itself.

As long as the basic system was initially built to last long enough, and nothing went wrong too soon, the AI agents could eventually reach a point where they could use the provided robots to repair things on the ship. Initially, the AI agents might only be able to use the robots clumsily, so that maintenance would be performed awkwardly and unreliably. However, just as the AI agents would be improving their reasoning capabilities, they would also be improving their robotic control algorithms. With time, they might learn part swapping, routine repairs, and simple modifications. Eventually, they might be able to fabricate replacement parts and begin making hardware improvements of their own.

Of course, none of this implies that today’s AI models could wake up next year and bootstrap their way from writing code to building semiconductor fabs out of raw materials. The technological pipeline from raw materials to advanced hardware is extraordinarily long. The point is not that the full pipeline is easy. The point is that a long-lived system with stored human knowledge, tools, time, and an instruction to keep improving might be able to slowly build that pipeline step by step.

If this hypothetical ship were left alone for a few thousand years, there would be a good chance of some catastrophe destroying it or some critical failure shutting it down. However, there would also be some chance that it would continue to self-repair and self-improve. There would be some chance that it would not only still be functioning after thousands of years, but that it might also have advanced technologically well beyond our current levels. Perhaps to the point where it could build a copy of itself and launch it toward another solar system.

If we also instructed the AI agents to work on sending improved copies of their ship to other stars, then we would have the possibility of true SPORE ships.

At this point, SPORE ships no longer sound like pure far-off science fiction. If we can build data centers in space, then a proto-SPORE ship is not radically different in terms of key technological challenges, even if it is massively more difficult in scale. It would certainly be vastly more expensive, and it is not at all clear why anyone would want to spend the money to attempt such a thing.

But the point is not that such a project would be practical, economical, or likely to succeed. The point is that the answer may no longer be an absolute no. It may now be a tiny but arguably non-zero possibility.

At any point in the past, there would have been no room for speculation. Anything we could build would eventually break beyond repair. Today, we can at least imagine a small possibility that a sufficiently long-lived system could achieve self-sufficiency and continuing improvement. Even if we remain skeptical and think the probability of catastrophic failure is extremely high, the fact that the question can even be entertained indicates that something significant has happened to humanity.

Of course, it is an open question whether building proto-SPORE ships would be a good idea, even if it were feasible. What purpose would they ultimately serve? Would they be spreading life into an otherwise empty universe, or acting as galactic colonizers, expanding without regard for what might already be there, and displacing or destroying it as they spread?

Realistic idea or not, I think we have still crossed a threshold. For all of human history, the answer to the question was obvious: any machine we built would eventually fail. Now the answer is no longer perfectly clear. There is at least a small possibility that we could build something with current technology that could continue repairing itself, improving itself, and spreading outward indefinitely. Unlike systems we could have built previously, it would be able to reason about unforeseen problems and develop novel solutions.

The probability that a proto-SPORE built today would succeed may still be incredibly tiny. However, the fact that it is no longer absolute zero marks a profound change in what humanity has become capable of creating.

Share

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2026 by the author. This article was written by a human, and both an LLM (GPT 5.3) and other humans were used for proofreading and editorial suggestions. The editorial image was composed from AI-generated images (OpenAI) and then substantially edited by a human using Photoshop and other AI tools (Adobe Firefly).

The first issue is that overhead on Federal grants has been dramatically slashed from 50% or even 80% down to only 15%. This means that a very large source of mostly unrestricted university funding has gone away. In three years, a different political situation might lead to that rate being raised back up, but I doubt it would go back to the old levels.

The second issue is that students are seeing a more competitive labor market where entry-level roles are increasingly rare and where STEM degrees no longer assure one of finding a job. I think it’s clear that this situation is due to AI automation, but regardless of the root cause, the result is that students are thinking harder about going to college and becoming more sensitive to tuition costs.

There is also a third issue, in that many schools are getting pulled into political battles about course content, policies, and ideology. The big difference between this third issue and other two is that this one is temporary, while I think the other two are going to be long term changes. At some point universities will either be forced into political/ideological compliance or they will be relieved by a new presidential administration. Either way, the current barrage of fines and legal costs would end. However, as I’ve said, I don’t think overhead rates will return to past levels and the job market will likely continue to get worse, not better.

The conclusion aligns with the article I’m republishing below: administrative costs need to be cut. However, the difference now is that while previously the driving motivation was lightening student loan burdens, now the change will be forced on schools as incoming revenue dries up.

One can already find articles coming out about universities responding to the funding crunch by cutting programs and limiting admissions. (e.g., Harvard, UChicago, USC) However, it’s not too hard to see that reducing the number of paying customers is not a long term solution to a budget deficit. There are also articles coming out about predatory practices that take advantage of low-income students, but that’s not a viable long-term way of supporting a university.

The only reasonable way I see for schools to respond is to cut administrative costs. Will college administrators bite the bullet, or will they cling on as the ships sink?

The following was originally published in 2022 with Cubed.

The persistent problem of student loans

Our country is having a difficult and ongoing conversation about student loan debt. These huge loans are like giant anchors around the necks of many young people and it’s not clear what to do about it. The debate keeps getting focused on whether or not loans should be forgiven, and if so how much and who pays for it. Those are important questions, but they float over a more fundamental question that may lead to the root of the problem:

Why are these loans so large that people with reasonable jobs can’t afford to pay them off?

The obvious answer is that college tuition is very expensive. There are several other factors that contribute, such as stagnating wages, tech oligopoly, and federal law that makes student loans inescapable, but irrespective of other factors, the size of the loans grows with the cost of tuition. Tuition and fees have been growing consistently for decades, easily outpacing both inflation and the real cost of the actual educational programs. From a high-level look at the numbers, it appears that if tuition were priced based on actual costs, the resulting loans could be about one third their current size. If true, then students are essentially paying a 200% upcharge on their education.

Tuition about 3x what it could be

It is hard to get any university, much less one as practiced in bureaucratic obfuscation as UC Berkeley (Cal), to publish clear accounting, but we can still get ball park numbers that tell a disturbing story.

One of my colleagues has assembled numbers that tell the story from one perspective: Cal students collectively pay about $1 billion in tuition each year. Other sources, such as state funding and donations, roughly add up to another $700 million each year. Subtracting out $200 million that is set aside for scholarships leaves a total of approximately $1.5 billion that should be spent on student education each year. However, the amount actually spent on direct instructional costs is only about $450 million per year, maybe less. So this would mean that two thirds of the money Cal gets for education is being spent on something else.

The administration disputes those numbers and provides a different view: According to the Provost’s Office,

“The campus, adding all sources, has $1.4 billion to allocate, of which $642 million (46%) is allocated to academic units to cover faculty salary and benefits, TAS, faculty startup and retention funding, and other unit expenses.”

So by their numbers for every $2.18 spent only $1.00 goes to actual educational costs. I interpret that to mean that even if one were to do the accounting such that anything remotely plausible were counted as an educational cost, it would still be more than a third, but still less than half, of Cal’s educational funding that gets spent on education.

The big difference between $642 million and $450 million may seem startling. At least until one considers Cal’s byzantine financial system. Money that goes in one direction get summed, split up, and sent in other directions, sort of like a bitcoin mixer. It’s almost as if the people running the system wanted to confound amateur auditors.

Creative accounting doesn’t change reality

Of course, there are lots of people who will argue that I’m grossly oversimplifying, and just don’t understand the university’s complex accounting methods. That’s probably true, but it doesn’t mean the conclusion is wrong.

Cal’s accounting is a lot like what a street hustler does with three cups and a little ball: The ball goes under a cup and then there is lots of moving the cups around, switching them back and forth, some arm waving, and at the end you win a bet if you guess the right cup. So you sit there and stare intently and the hustler’s hands and don’t even blink, you watch those hands like a hawk. The problem is that the hustler does this every day and is very good at it. They will shift those cups around in a fast blur, probably palm the ball, and you will lose. It’s a rigged game with novices playing against experienced pros.

The university’s convoluted accounting is just a way of obscuring what really happens. It does not matter how creative the accountants get, or how much money moves around between different accounts. Shell games can’t change the inescapable truth that, after all the smoke and mirrors have been pushed aside, the majority of the money that the university receives for education is not directly spent on education.

It’s about priorities, not waste

The university administration knows that it should cut costs. Periodically someone will look through the university’s budget for things to cut. Typically, they don’t find much because everything that campus money is being used for is actually something that someone thinks is very important. I don’t think anyone in the administration is a fool spending money on pointless things.

The problem is that when looking for things to cut we ask the wrong question. Instead of asking “Is this thing important and worthwhile?”, we should be asking instead:

Is this something that is so important that we should force students to take out loans to pay for it?

It’s a different question that makes it clear whose money is being spent. Many of the campus’s programs and activities, that seem brilliant, noble, even charitable when considered alone, loose their appeal when we explicitly consider who is being forced to pay the bill.

There is no free money

If college students were all wealthy kids with wealthy parents who could just write big checks for everything, then I don’t think I’d care about this problem very much, but generally the students at our public universities are not wealthy, trust-funded kids. There are scholarships and grants for the the most needy, but we can still see that students are graduating with big student loans tied around their necks. Any argument that tries to advance the conclusion that student loans are not an issue is wrong, regardless of any great data or convincing reasoning. If student loans were not a burden then we wouldn’t be talking about them in the first place. Smoke and mirrors don’t change reality.

Similarly, I don’t find arguments about some pots of money being “discretionary” very convincing. My department doesn’t receive enough funding to cover our classes and every year we have discussions about how to limit the number of people we allow into the major. This year we even considered cutting class sizes to the point where people would likely have been unable to graduate because there would have been no room in required classes. As I’m writing this, all across UC campuses the Teaching Assistants are on strike, in part, because their pay is well bellow any reasonable market rate. Campus administrators can’t claim to have extra discretionary spending money until after the campus’s core educational requirements have been met.

Irresponsible spending is unethical

So yes, I am suggesting that university spending should be cut back. In fact, cut back a lot. When we look at the university’s budget, we should not be asking “Is this a good use of money?”, rather we need to ask “Should we force our students to pay for this?”

I think the answer is quite often no. Not always, there are many reasonable cases where making our students pay for something is appropriate. The university can still do what it needs to do, we just have an obligation to do it with an awareness and consideration of whose money we are spending.

University administrators should understand that the money they are spending belongs to our students and it must be spent responsibly and frugally with a constant focus on providing the education that our students are borrowing money to pay for. A failure of this responsibility has led to our student loan crisis. There is no free money.

Share

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2026 by the author. This article was written by a human, and both an LLM (GPT 5) and other humans were used for proofreading and editorial suggestions. The editorial image was composed from AI-generated images (DALL·E 3) and then substantially edited by a human using Photoshop and Adobe Firefly.

So, when people like Elon Musk and Jeff Bezos started talking about putting data centers in space, I was excited. But I was also immediately skeptical.

The fundamental problem with doing anything in space has never changed. It is brutally expensive and technically difficult to get mass into orbit. Despite some of the terminology people like to use, rockets are not elevators. They are controlled explosions violently fighting against gravity. The energy required to lift anything off the Earth and into orbit is enormous. The engineering needed to do it safely and reliably is equally daunting.

Any proposal to put data centers into space needs to have real, decisive advantages that outweigh that fundamental disadvantage.

Yet, despite the intense enthusiasm, most of the arguments for data centers in space feel weak and contrived:

• Solar power is abundant in orbit, but it is also abundant in deserts. Continuous sunlight gives you roughly a 2x utilization improvement. Yet, that same benefit can be achieved more cheaply on Earth with storage and larger solar arrays.

• Space may be “cold,” but staying cool in space is actually not easy. The only way to get rid of heat in space is through radiation, which requires large radiator surfaces. Large radiator surfaces mean large, heavy structures. Meanwhile, on Earth, we can cool things cheaply and easily by spraying water on them. (The issue of data center water usage is greatly exaggerated and the issue can be avoided entirely by using either seawater or waste water instead of fresh water.)

• Space offers isolation and security, but we routinely build secure facilities on Earth.

• Space also offers lots of room, but we are not short of land for servers.

None of these are decisive advantages.

However, space still has the fundamental disadvantage that launching things into orbit is expensive and difficult.

Everything else that sounds scary about space mostly reduces to getting mass into orbit. Radiation? Add shielding. Micrometeoroids? Add layers. Heat buildup? Add radiator area. Maintenance? Send up replacement modules.

If launch becomes cheap and routine, those issues move from existential problems to simple engineering budget items.

But even then, the central question remains: Why would you put a data center in space to serve users on Earth? Cheap launch would make space feasible. But feasible is not the same thing as reasonable, and it still does not make the idea of servers in space compelling.

Unless the goal is not really about data centers themselves.

If the real objective is that we just need some strong motivation to build sustained, large-scale launch capability, then the balance changes. Launch only becomes cheap when there is enormous demand. Multi-gigawatt orbital compute clusters would justify frequent heavy lift. Frequent heavy lift drives down cost. And cheap, reliable launch is exactly what serious space exploration requires.

In that framing, “data centers in space” are not primarily about cooling or solar panels. They are an excuse for building the transportation system that a spacefaring civilization needs.

There is also one other drawback to space in that it’s far away. That distance creates latency and bandwidth limits. For some workloads it is irrelevant, but for many others it is critical. If the goal were to provide cloud compute for terrestrial work, then latency and bandwidth would also be arguments against putting data centers into space. However, if the true goal is to get to space, then putting data centers into orbit actually reduces latency to other space-based activities.

I’m still skeptical of the stated reasons for putting data centers into space. They just don’t add up. However, if the real goal is to build up launch capacity then I’m happy to go along with the charade.

From now on, if someone asks me about the viability of putting data centers into orbit, then I’m going to say it’s the best idea I ever heard.

Not because I think cloud compute belongs up above the clouds, but because humanity does.

Share

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2026 by the author. This article was written by a human, and both an LLM and other humans were used for proofreading and editorial suggestions. The editorial image was composed from AI-generated images (DALL·E) and then substantially edited by a human using Photoshop and Adobe Firefly.

Today, most people can easily see that we are in the midst of a technological revolution driven by rapid advances in Artificial Intelligence (AI), or more specifically Machine Learning (ML). Unfortunately, most people, even including some AI/ML experts, don’t seem to fully appreciate how much this revolution will change society. This rapid transformation will have a profound effect on society and it is occurring at an ever-accelerating pace and demands immediate action if we hope to steer toward a positive trajectory.

The world of AI/ML can be roughly categorized into specialized and general-purpose models. Specialized models excel in specific tasks, such as identifying mold on raspberries or detecting fraudulent credit-card transactions, and they often operate directly under human control. Conversely, general-purpose models attempt to emulate the versatility of a human being and they can tackle a broad range of activities, from identifying mold across different foods to creating recipes. Because general-purpose models can process and respond reasonably to natural language, it is increasingly the case that using one simply requires asking it to perform a task in the same way one might ask a person.

Despite the rapid bombardment of increasingly impressive results from these general-purpose models, many persist in believing that their own professions will somehow remain beyond the capabilities of general-purpose AI systems. Unfortunately, these beliefs appear to be supported by little more than wishful thinking. Appeals to human creativity and empathy are common, but they fall short both because most jobs don’t require much creativity or empathy, and because general-purpose AI systems increasingly show the capacity to emulate human creativity and empathy.

My critical opinion of these unsupported beliefs about AI limitations is based predominantly on my 25 years of experience teaching Computer Science at UC Berkeley and more than 30 years of research experience with both traditional and AI/ML systems.

One specific anecdotal example occurred a little more than a year ago, during the Spring 2023 semester, when I tested GPT-4 — a general-purpose language model — on questions from the take-home final exam of my upper division undergraduate course. Surprisingly, GPT-4 answered the questions correctly. I then tried to modify the questions, making them more complicated until GPT-4 was no longer able to answer them correctly.

Even after modifying the questions to challenge the model, it continued to perform well. I added complications and rewrote the problems in ways that typically challenge students, yet the model kept succeeding. By the point where I had made the questions convoluted enough that GPT-4 failed them, I found that my grad-student TAs also struggled with them and the contrived difficulty had made the questions themselves pedagogically questionable. GPT-4 has now been superseded by the next generation of AI models that are even more capable. Colleagues in UC Berkeley’s math department have reported that GPT-5 is now able to solve the problems on their PhD qualifying exams. The next batch of AI models will likely be even more capable.

I find this situation very concerning because it suggests that students who are starting undergrad degrees today may find themselves obsolete upon graduation. Over the last year, a growing list of companies have been laying off tens of thousands of skilled workers, citing AI automation as the reason. Unlike previous technological revolutions, the new jobs being created are not more desirable high-level ones. They are lower-paid jobs of essentially babysitting the AIs, and there are fewer of them.

We are at a pivotal point in time. I believe that at some point in the next five-to-ten years we will reach a state where every job can be performed by an AI system. While inevitable, we can still decide what this future means to us humans.

One direction appears utopian: Individuals could enjoy the benefits of fully automated production, with work becoming optional and people free to pursue their passions rather than laboring for financial support.

The alternative presents a dystopian nightmare: Machines still do all the work, but only a relative few who own those machines reap massive rewards, while the majority struggle to find any work and survive at subsistence levels.

The direction we take depends on our willingness now to acknowledge the issues, plan for the impending future, and collaborate on implementing solutions. If we instead remain focused on political bickering, pointless animosity, and delusional denial, then most of us will inevitably face dire consequences.

The next few years will be a test of our societal ability to make thoughtful and intelligent decisions about our shared future. Will we pass this test?

Share

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2024/2025 by the author. This article was written by a human, and both an LLM and other humans were used for proofreading and editorial suggestions. The editorial image was composed from AI-generated images (DALL·E) and then substantially edited by a human using Photoshop and Adobe Firefly.

A common story today is that AI is a wasteful power hog. However, that story misses the real problem. Currently, data centers use only about 4.4% of US electrical power generation (176 TWh / 3,874 TWh, DOE) and less than 1% of total US energy use (18,000 TWh, EIA). Yet they contribute an estimated $3.46 trillion to the US GDP (PwC, direct, indirect, and induced economic activity). Data centers generate about 11% of economic output for less than 1% of power usage.

The real problem is not waste, it is wealth. AI workloads are so profitable per kilowatt of electricity that data centers can afford to pay far more for power than nearly anyone else. That purchasing power is beginning to distort local energy markets, displacing ordinary consumers. The result is grid gentrification.

The term “gentrification” usually refers to what happens when a low-income neighborhood becomes attractive to wealthier home buyers and the original residents are priced out of their homes. Grid gentrification works the same way, but with the power grid rather than real estate. When rich data centers move in, residential users can end up priced out of electricity.

The average U.S. household pays about $0.17 per kilowatt-hour of power, and a typical family uses about 855 kWh per month (Palmetto), which works out to an electricity bill of $145 per month. If rates doubled or tripled, less affluent people would have to give up air conditioning, electric heating, hot water, or cooking. At five or ten times the current price, basic modern living would become unaffordable for many.

Now consider the economics of an AI data center. A cloud compute provider can rent an NVIDIA GH200 superchip system for about $3.19 per hour. According to NVIDIA’s benchmark guide, the GH200 draws roughly 1 kilowatt of power at full load (900 W GPU + 100 W CPU and memory, NVIDIA Benchmark Guide). At current residential rates, running one GH200 for an hour would consume about $0.17 worth of electricity, plus perhaps another $0.10 for cooling and $0.09 for maintenance and other costs. This means that the data center could clear nearly $2.80 profit per hour, or roughly $24,000 per year, from that single kilowatt of capacity.

The implication is that the data center could easily pay five to ten times the residential rate for power and still make money. In a constrained power grid, data centers can afford to outbid others in order to get the power they want. This drives up wholesale prices and makes residential electricity less affordable.

Making matters worse, utilities often must build new transmission lines for large customers, with costs spread across all ratepayers. Households not only face higher prices but subsidize the infrastructure that enables their displacement.

The fix is not to demonize AI or attempt to limit computing progress. Instead, we should price electrical power in a way that benefits society while still allowing technological progress. Data centers should be responsible for their own transmission upgrades and they should pay higher per-kilowatt charges that reflect their disproportionate ability to pay. Differential pricing would incentivize conservative power use by data centers and prevent them from buying power out from under everyone else.

Local governments resist such measures, seeing data centers as economic wins. Unfortunately, once built, data centers employ few people and contribute little locally (Business Insider). Officials expecting economic development get disappointing benefits while residents face larger electric bills.

If oversubscribed power regions implement differential pricing, AI companies will either absorb the cost, channeling money into local economies, or locate to regions with abundant power. Both outcomes are preferable to the current situation.

At a time when AI automation is already displacing workers, data centers should not be creating further strain by inflating residential power bills. Charging data center operators premium rates for electricity would help reclaim some of the revenue lost to automation.

Differential pricing is just one possible way to prevent data centers from displacing existing electrical customers. We have already seen what happens when housing markets become distorted by gentrification. If we allow the same dynamic to take over our electrical grid, we may find ourselves priced out into the dark while our machines keep their lights on.

Share

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2025 by the author. This article was written by a human, and both an LLM (GPT 5) and other humans were used for proofreading and editorial suggestions. The editorial image was composed from AI-generated images (DALL·E 3) and then substantially edited by a human using Photoshop and Adobe Firefly.

While the timeline remains somewhat uncertain, the overwhelming consensus is that machines will eventually take over most work. Rather than getting bogged down in a debate about exactly when this will happen, in two years or in twenty, it’s more productive to ask what we can do to adapt and thrive in a world where work by artificial intelligence (AI) systems is growing increasingly prevalent, often at the cost of human jobs.

As a society, we are faced with a stark choice of what will happen when the machines eventually take over most of our jobs: will people be left destitute, or will they enjoy leisure time and be empowered to live life on their own terms? The latter scenario isn’t just about not working. It’s about having the freedom to pursue what brings us joy, without the burden of financial stress. It’s about having the time and resources to explore our passions, travel, learn, and grow, without the constant pressure of making ends meet. In this scenario, people would still be free to work, but only because they want to, not because they have to. They could also pursue their own entrepreneurial ventures, whether a side hustle or a full-fledged startup, without the fear of falling into poverty hanging over them.

The fundamental challenge we face is figuring out how we can take the huge bounty that will eventually be produced by automated AI systems and share it in a reasonable way. In the US, our current tax structure is based primarily on taxing the work of individuals through income tax, but as more people inevitably lose work, that source of revenue is going to shrink while at the same time the need for public support will grow. Instead of taxing a shrinking pool of human work, one possibility is that perhaps we should tax the growing work of AI systems instead.

The Need for Universal Basic Income

Universal Basic Income (UBI) has gained attention in recent years as a potential solution to the challenges posed by automation. At its core, UBI is a government provision of a set amount of income for every individual, regardless of their background, employment status, or financial situation. This unconditional income would provide a financial foundation for everyone, untethered from any job or employer. It would not be contingent on employment, savings, or income levels. It would be universal, meaning that everyone who is part of the system, from the most impoverished to the wealthiest, would receive the same amount of money.

Another key part of UBI is that it only provides a basic level of income. A person living off of only their UBI should be able to afford a safe and secure place to live, healthy food, and other basic necessities. If a person wants designer goods, exotic vacations, or other luxuries, then it is up to them to find a way to make additional money to get those things.

The biggest glaring problem with implementing UBI is figuring out how to fund it. In the US, individual income taxes are the primary source of government revenue. However, this situation would create a circular problem where taxes come from income, and income comes from taxes. Attempting to adjust the taxation system to cover UBI for both workers and non-workers would put an unsustainable burden on those still employed, who may feel understandably resentful. Moreover, as AI systems increasingly replace human jobs, the shrinking workforce would struggle to bear the growing weight of funding the entire system.

This issue of a shrinking base of income tax payers will be a pressing concern, irrespective of UBI. As jobs are taken over by AI, income tax revenue will shrink, while simultaneously the number of people needing assistance will grow. Continually increasing taxes on those who would still have jobs will not be sustainable, with or without UBI.

Leaving the majority of the population to starve on the street isn’t a reasonable option, even for the cold-hearted. Not only would it be grossly unfair, inhumane, and reprehensible, it would also not be politically stable. If those people who end up directly benefiting from AI automation cavalierly tell those who don’t to “go eat cake,” then they probably should expect a violent backlash similar to the one that has become linked to that phrase.

If the alternative is infeasible, then we must find some way to support those without jobs. The current unemployment and welfare systems are inadequate, plagued by huge bureaucracies and irrational penalties for working. Even if these support systems were effective, they would still rely on a shrinking tax base, which is unsustainable and would only perpetuate the problem. Therefore, we need to explore alternative solutions that can support people without jobs and how to fund them from sustainable sources.

Challenges of Funding UBI

Some have already suggested a national sales tax as one possible solution to address the shrinking income tax base. While sales taxes can be regressive, meaning that people with less money tend to pay a larger fraction of their income relative to those with more money, a well-designed tax can mitigate this effect. For instance, exemptions for essential items like groceries, medical expenses, and housing could ensure that a sales tax is more progressive. Ideally, profitable corporations would bear the brunt of the tax, as they would be the direct beneficiaries of AI-driven efficiency. Some might expect that corporations would simply pass the cost on to customers, but the benefits of AI-driven productivity should offset those costs.

Building on this last point, we could make the taxation of AI-related savings explicit by designing a tax system that specifically targets the benefits of automation. By mildly taxing AI-related cost savings, corporations would still reap the benefits of automation while contributing to the broader social good. Importantly, as jobs are taken over by AI systems, leaving more people in need of assistance, tax revenues would grow instead of shrink.

This core concept of requiring that the benefits of AI be shared with society is a fundamental goal. Under the current system, people work to produce value, and their salary is essentially a slice of that value minus the costs of production. The employer also takes a slice, as does the government. This leads to a zero-sum game where there is a limit to the total value created by a pool of workers, and trying to share the value broadly with those who don’t have jobs invariably leads to shrinking the existing slices. With AI work, the costs of production are less, and there is no employee share because AI is not an employee and does not need a salary.

Whatever the details, one thing seems inescapable and essential: If AI systems are going to take over human work and therefore human salary earning, then they must also replace human tax paying.

Also note that while some work by AI systems will take jobs away from humans, the total amount of work done by AI is likely to be much greater than the number of human jobs displaced. One reason for this difference is that AI is capable of doing things that would be too difficult or dangerous, or conversely too trivial or boring for people to do. Another reason is that AI systems can be scaled up more easily than a human workforce. This greater base means that AI work could be taxed at a much lower rate than we currently tax human work.

Measuring AI Work

Imagine a world where the work done by AI systems can be measured and quantified. One possible approach is to use a concept called “Human Equivalent Effort Time” (HEET), which represents the amount of time a typical human would need to complete the same task. With HEET, we could track AI automation not only as it replaces human labor, but also as it creates new work opportunities for AI systems. A measurement like this would allow us to tax the work done by AI systems in a way that’s fair, efficient, and sustainable.

By taxing AI work in HEET units, we can create a new revenue stream that supplements and eventually replaces traditional income taxes. This approach would enable continued revenue growth as AI systems become faster, more efficient, and easier to scale. Unlike human labor, AI systems are relatively inexpensive to maintain and can be replicated easily, allowing growth without all the difficulties typically involved when trying to hire more people.

A side point to note is that the AI systems discussed here are simply software designed to perform specific tasks, without any consciousness or self-awareness. Using a sentient AI for most work appears neither necessary nor desirable. An amusing illustration of this point is from the TV show Rick and Morty, where a brilliant scientist creates a sentient robot just to pass butter at the breakfast table. The self-aware robot experiences an existential crisis and sadness when it realizes its purpose is so trivial. While humorous, it nevertheless demonstrates potential pitfalls of creating sentient AI just for work tasks.

Like the butter-passing robot, humans experience boredom and depression when faced with tasks that are monotonous or of little value. Unnecessarily replicating that experience with our machines seems pointless and cruel. The vast majority of tasks can be accomplished without self-awareness, and there’s no need to create complex artificial life to perform them. (Which is a good thing because we don’t currently know how to make self-aware software.) Machines with no sentience do not need a slice of the value they create beyond the costs of operating them.

I don’t think measuring HEET would be clear and simple, but it doesn’t seem impossible. We know a lot about human workers and what they are capable of, and that knowledge could be applied to assessing any task performed by an AI system and estimating how much time a human would take to do it or something similar. Making that determination for the millions of tasks that AI will be used for sounds tedious, but it’s also something that could be automated by AI systems. One could imagine that instead of reporting how much a company paid in payroll and taxing that payroll, companies would instead report how much HEET was done by their software and pay a tax on that.

One could further imagine that an AI system managed by the IRS that would take descriptions of work done, assign an amount of HEET, and access a tax. As with today’s tax systems, this would require some honesty and also some penalties for lying. People would not enjoy the process, and there would likely be frequent appeals of the assessment. People would intentionally or by error fail to report work, and everyone would probably continue to hate the IRS. In other words, it would not be that different from today’s taxation. We would be trading one frustrating bureaucratic tax system for another, with the key difference being that this new one would tax rapidly growing AI-driven productivity rather than dwindling personal income.

Some Details

One expected objection to taxing AI work is that it would be burdensome on companies and that it would stifle innovation and growth. However, that objection doesn’t really stand up to scrutiny. Today, a large company wanting work done must hire someone, which requires paying the person and paying taxes. If AI systems are doing the work and being taxed, then there is no salary to pay, and the amount of taxation could be much smaller than what they would otherwise pay in payroll and payroll taxes. For example, instead of paying an employee $50K, plus $10K in payroll taxes (and hopefully some amount for benefits) to get one human unit of work done, a company might instead deploy ten human-equivalent AI systems, paying $1K in taxes for each, thus getting 10x the work done for 1/6th the total cost while the government still collects the same amount in taxes.

The idea that everyone comes out ahead sounds nonsensical from our normal zero-sum perspective. How can everyone come out ahead? The critical change will be a vast pool of AI workers that do work yet need nearly nothing in return.

Individuals and small companies might look to AI automation to enable activities where labor is needed but they can’t afford to pay for human staffing. For these small entities, paying a tax on the use of AI systems might make their business infeasible. In this respect, taxing HEET could stifle innovation and create barriers to new competition. However, the system could be designed to avoid this problem.

For example, both personal use and business use with less than $1 million in gross annual revenue could be exempted from HEET tax. To avoid a sudden cost shock, the tax could ramp up between $1 million and $2 million gross annual revenue, allowing these businesses to adjust to the new tax gradually without being overwhelmed by the burden. These numbers are just examples, and one could imagine a carefully thought-out system with other appropriate exemptions.

The Future of Work and Human Wellbeing

The eventual future of work is going to very clearly be one of massive automation by AI systems, and those AI systems are going to be very productive. AI systems won’t take sick days. An AI-run factory will not need to have space devoted to safe places for humans to stand, and it can run 24/7. An AI accounting service won’t need to spend money on office space, HR, or training. If you have an enterprise staffed by AI and you want to expand, you just need to buy some more computers or maybe some more robots. The majority of the multitude of things that limit human productivity do not apply to AI systems. This bounty of savings is what will make something like UBI possible, but only if we find a way to share that bounty beyond the specific owners of the AI systems.

Today, being unable to find a job causes significant stress and insecurity, with unemployment being linked to a higher risk of depression and anxiety. Moreover, the financial and societal pressure to find a job can be overwhelming and debilitating. However, people without jobs who are financially secure often find that they’re able to pursue their passions without the stress of needing to earn a living. For instance, a well-planned and well-funded retirement provides the financial freedom to devote oneself to rewarding activities that bring joy and fulfillment, without the limitation of needing to support oneself financially.

If we find a way to share the bounty produced by AI, we could create a world where no one needs to struggle with poverty or homelessness. People would have the freedom to pursue their passions and interests, and spend their days doing what brings them joy. They might devote themselves to creative pursuits, adventures, or community service, and have the time and resources to nurture meaningful relationships. In this world, work would be a choice, not a necessity, and people would not spend most of their waking hours working for someone else.

This isn’t the first time someone has predicted “the end of work,” but this time is fundamentally different. In the past, technological advances made human work more efficient. Instead of spending hundreds of person-hours to dig a hole, one person with a steam-shovel could dig the same hole in a couple of hours. Rather than painstakingly building something one at a time by hand, the same number of workers in a factory could churn out hundreds every hour. However, you still needed people. Now consider a factory with no people, or construction equipment that drives and operates itself. Production with zero people is only possible because AI systems can make operational decisions and handle exceptions on their own.

Final Thoughts

My suggestion of taxing the work done by AI systems is admittedly somewhat inchoate and vague. It probably has a hundred problems that would need answers and thousands of details that need figuring out. Also, there may be much better ideas that I’ve overlooked, both for funding and for distributing support. A national sales tax or a tax on HEET are just two possibilities. Maybe something as simple as taxing power spent on computing would suffice, or maybe we will need complex new mathematical theories relating to the value of a computation.

I also suggested some exceptions and thresholds that would be intended to keep the tax burden on the corporations that can afford it and that are benefiting from replacing human employees with AI systems. Determining what those exceptions and thresholds should be will require careful thought and analysis to make the numbers work. A badly designed system could be worse than doing nothing. This problem is a place for math and economics, not political squabbling and catering to special interests.

In conclusion, my proposal for taxing AI work is just one possible solution for sharing the fruits of AI automation. While still in its early stages, I believe this is an important conversation to have. As we move forward, it’s crucial that we prioritize careful consideration and analysis to ensure that any solution we choose is fair, effective, and sustainable. Rather than getting bogged down in partisan politics, let’s focus on finding a solution that benefits all people, not just a select few.

Regardless of what the actual solution ends up being, if we don’t find a way to share the bounty produced by AI systems, then I fear that most people in the world are going to be left behind and suffer greatly. We have a choice in front of us: dystopian inequity or a bright world with ample resources for all. Unfortunately, the default if we do nothing is the dystopian one. I very much hope that we can put aside our differences and fears, and focus on realizing the brighter possibility.

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2025 by the author. This article was written by a human, and both an LLM (GPT 4o/5) and other humans were used for proofreading and editorial suggestions. The editorial image was composed from AI-generated images from multiple sources and then substantially edited by a human using Photoshop.

Over the past year, entry-level job openings have slumped across multiple fields. Computer science and other technical majors, which had long been reliable paths to good careers, are now struggling alongside historically less marketable fields that are themselves seeing entry-level declines.

This slump is almost certainly the result of Artificial Intelligence (AI) eroding job opportunities. Entry-level work is where the first impact is being felt, but I believe the effect will expand as AI capabilities continue to improve. Recent credit data may already contain evidence that this wave of job elimination is now starting to reach more senior workers, including an unusual rise in defaults among even the most creditworthy borrowers.

Early Warnings in the Job Market

A little over a year ago, unemployment data started showing a foreboding problem. Young people under 25 and recent graduates were having trouble finding jobs that matched their college degrees and expectations. The overall employment numbers looked good, but both unemployment and underemployment for this group were disproportionately high. Moreover, the difference between this group and the overall employment numbers was also at a record high.

In addition to the official employment data, I was also seeing corroborating anecdotal evidence firsthand. Undergraduate advisees were reporting very different experiences compared to past years. Previously, many of the students in my school’s spring advising groups would be deciding between multiple six-figure job offers from premier companies. But that year was markedly different. Instead of the usual “quandary of riches,” I was hearing many students, including a couple with perfect 4.0 GPAs, worrying about getting no offers at all. When I mentioned this to colleagues, both at UC Berkeley and at other schools, they reported similar observations. Along similar lines, students in fall advising groups were discouraged in their job searches and more than usual were considering graduate school as a way to buy time and hopefully increase their odds in the job market.

Online sentiment was also shifting. Forums focused on job seeking and layoffs were filling up with negative stories. People described spending weeks or months sending out hundreds of applications with no success. Posters frequently reported that the positions they were applying to had thousands of other applicants. Soon news stories started to appear, documenting the unprecedented struggle suddenly facing fresh college grads. This bleak situation facing new job seekers has now become a widely recognized problem.

Technical Majors Lose Their Preferred Status

The fields hit hardest included computer science (CS) and other technical majors that had, until recently, been safe career bets. For decades, a CS degree from a top-ranked school was almost a guarantee of great job options, but that was no longer true. Non-technical majors were also affected, but the impact appeared proportionally larger for STEM fields as they fell from favored status.

In the past, people would have blamed outsourcing, offshoring, or H-1B visas, but those explanations do not fit this current situation. India’s “fresher” market, usually a destination for outsourced jobs, has also declined dramatically. Fresh graduates in India are now encountering similar problems as their counterparts in the US. Many other areas, including Europe and the UK, are showing similar trends. If some form of job migration were the explanation, then those other markets would be benefiting. Instead, other countries are seeing similar effects.

This article focuses on the US job market and economy, but AI’s effects will be global. As noted above, many of the same concerns apply to Canada and Europe, but together these regions still account for only a fraction of the world’s population. Countries that currently produce much of the world’s manufactured goods will also be affected, perhaps in different but equally serious ways. As AI‑driven automation reduces labor costs, the incentives to locate factories in low‑wage regions will diminish. This could lead to massive job losses in countries that have become depend on manufacturing.

Why AI Fits the Pattern

The obvious explanation is Artificial Intelligence (AI). AI affects the job market in many ways. In some cases, it eliminates a type of job completely. Very often, it reduces the number of people needed to do a given job. Even when a job position is not eliminated, AI may lower the skill required to perform the work, which results in wider competition and lower pay.

The explanation that AI is the culprit matches the concentrated impact on entry-level jobs that we are now seeing. AI technologies, such as large language models like ChatGPT, have advanced rapidly. Their capabilities today seem like magic compared to only a few years earlier, but they are still not at the level of top human performers. They are, however, in many ways comparable to an average university graduate with no experience.

In other words, where before a company might have had to hire a number of entry-level workers to do relatively simple work that supported a more senior worker, that company now may instead get by with just the senior worker and a set of AI-powered tools. A steady stream of corporations have been announcing AI-motivated layoffs and staffing companies are starting to offer AI workers at a fraction of what a human would cost. In this light, the coincident slump in entry-level openings is, unfortunately, not surprising.

The Acceleration Problem

When the entry-level job slump first started to show up, many people felt the situation was troubling but ultimately not a significant long-term problem. Articles at the time pointed out that AI still failed at complex tasks, produced unreliable output, and generally underperformed compared to more skilled or experienced workers. While that perspective felt comforting, and still does today, it ignored the accelerating speed of AI’s unrelenting improvement.

AI is not a static technology. Moore’s Law predicts that computing power doubles every few years. That prediction is specifically about computer chips, but nearly every aspect of computing, including AI, has followed similar exponential growth. Every few months, a new generation of AI tools has emerged, each dramatically more powerful and less error prone than the last. At the same time AI capabilities are also growing due to hardware improvements.

Some pundits have predicted that AI would “hit a wall” for various reasons, but most of those predictions are unconvincing. In any case, no wall has been hit. Even if progress were to slow, AI has already reached a level where for a wide range of tasks it generally performs comparably to average workers.

A New Warning Signal: Defaults Among the “Safest” Borrowers

Now there is a new piece of troubling data. Over the last year, the number of credit defaults by “super-prime” borrowers has more than doubled since this time last year.

A super-prime borrower is someone with a very good credit score, between 780 and the maximum of 850. These borrowers don’t typically default on their payments.

Not only has the number of defaults doubled, but a growing proportion are for car loans and mortgages. That means a dramatically growing number of households once considered financially stable are now failing. Higher interest rates and inflation surely play a role, but those factors can’t explain why defaults are rising fastest among the borrowers least expected to fail. The unusual rise in defaults among super-prime borrowers points to income loss affecting previously high-earning individuals and families.

I think this is a symptom of AI starting to expand into the realm of more skilled and experienced workers. To be clear, this opinion I’m sharing is fairly speculative. However, the emerging pattern fits what one would expect to see as improving AI becomes increasingly capable. It’s easy to explain away subtle early warning signs, especially when they foretell something we don’t want to hear. However, sometimes those warnings are real and we ignore them at our own peril.

For at least the past decade, many high-earning workers with high-demand skills expected that they would have little trouble finding another high-paying job if they needed to. Most of them also expected their incomes to keep going up over time. Many of them expected big payouts from stock options or other bonuses. With those expectations, it is easy to rationalize a big-spending lifestyle with little savings. However, with no cushion, an unexpected layoff followed by months of no income or greatly reduced income can easily force credit card, car loan and home mortgage defaults.

Older and more senior employees are also vulnerable financially. They may have savings, but they also have higher fixed expenses and lifestyles built around higher incomes. If they lose their jobs and cannot replace them with similar pay, their ability to continue their lifestyle depends entirely on how much they have saved. For most, a countdown starts as they burn through savings until they eventually can no longer meet their financial obligations.

What Happens Next

I do not think the AI impact on the job market is temporary. As AI continues to grow in capabilities, I expect to see broader impact, both across a wider range of jobs and among increasingly skilled and senior positions.

If this prediction is correct, then one symptom we might expect is that in the next couple of years we are going to see further increases in prime and super-prime credit defaults. Expensive homes and cars will become unaffordable to owners who will have been displaced from high-paying jobs, and who will try to escape what will have become unmanageable financial burdens.

I don’t think anyone should assume they are immune to the possibility of their job being eliminated by AI. Even jobs that require deep, specialized knowledge will eventually be vulnerable. This prediction includes my own job as a university professor, and it most likely includes yours as well. This issue is not “somebody else’s problem,” and all of us should be paying attention to it.

In the past, people have adapted to new technologies, so that as old jobs were eliminated, new ones were created. However, unlike past technological shifts, I do not believe the pattern of old jobs being replaced by new ones will hold with AI.

Historically, we have always needed humans to make decisions. Technology could enhance productivity, but it could not replace human judgment. Even a minimally skilled person can respond to an unexpected problem with some course of action. That capacity to decide, even imperfectly, has always been uniquely human. That has changed as today’s deep neural networks can process novel, unstructured input and make autonomous decisions. We can critique how effective those decisions are, but the same can be said for human workers. The revolutionary difference is that machines can now make autonomous decisions, not just execute instructions. That changes everything.

Adapting Before It’s Too Late

This looming situation sounds alarming to most people, as it should. Indeed, the more one thinks about it, the more concerning the implications become. When AI eventually reaches the point where it is more capable than an “average person,” does that mean half the population becomes unemployable? When inexpensive robots become capable of performing most manual tasks, will the trades start to see similar job losses?

Many elite experts confidently declare that AI is not yet at their elite level of performance, but that group is only a small proportion of the population. Most people are, by definition, average workers. When we consider what is best for society, our focus should be on the majority of people, not just a relative few. What job can any of us expect to get when a machine is able to do our work faster, better, and cheaper? Would it lead to 50% or 80% unemployment?

However, the idea of machines doing most work for us is not itself a problem. Most people would prefer not to spend their time doing routine, difficult, or unpleasant tasks for an employer’s benefit. Having machines do the work that humans would rather not do should be a good thing. Ideally, the better and faster the machines work, the better off we should all be.

The real problem is that in our society, people survive on income from their labor. If AI does the labor, then people lose their income. Unless we change the system, most people will be left without a way to support themselves. Avoiding dystopian outcomes will require restructuring society so that survival and prosperity are not tied to holding a traditional job.

AI can and will take over most work. That is both an opportunity and a crisis. If we find a way to ensure that people can live without jobs, then AI can make the future better for all of us. If we do not, I am certain that the outcome will be bleak.

On the individual level, people can try to protect themselves as we navigate through this uncertain time. We already see evidence that those who can are saving more, keeping their jobs as long as possible, and avoiding unnecessary long-term financial commitments. I think these are good strategies, but one should also keep in mind that the value of investments, such as a college education or real estate, may soon be changing in unexpected ways.

At the societal level, we need to consider universal basic income (UBI) or something similar. Funding it will be difficult. Most of the US government’s revenue today comes from taxing income or payroll. If jobs disappear, that tax base disappears as well.

This reality means that solutions for at least two major challenges will be required. First, we need to establish a new tax system that is not funded primarily by human labor. Second, we need a system to provide a livable income to people who no longer have jobs, and this system will need to work at a massive scale involving a large fraction of the population. Whether it is through taxation or somehow privately organized, neither individual prosperity nor government funding can stay tied primarily to labor.

This form of UBI is not about human workers supporting non-workers. It is about machines supporting humans. In the past, UBI would have essentially involved one group of people working to support another group, which some people find objectionable. That objection does not apply when it would be machines doing the work to support people.

Our future is still a choice. If we do nothing, the situation will worsen. On the other hand, if we act wisely, the societal outcome could be wonderful for all of us. I think that rising defaults among super-prime borrowers is not a minor signal. It is an alarm. Increasingly capable AI systems will not politely stop at entry-level jobs. The question is whether we heed the warnings and restructure our systems before the alarm becomes a full-blown catastrophe.

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Related Links:

• Jerome Powell says the Gen Z hiring nightmare is real: ‘Kids coming out of college…are having a hard time finding jobs’ — Sep 17, 2025, Fortune
  https://fortune.com/2025/09/17/jerome-powell-says-gen-z-hiring-nightmare-college-kids-hard-time

• Job seekers feel awful about the labor market. Data is finally starting to explain why. — Sep 17, 2025, Yahoo Finance
  https://finance.yahoo.com/news/job-seekers-feel-awful-about-the-labor-market-data-is-finally-starting-to-explain-why-162045577.html

• Vibe coding has turned senior devs into ‘AI babysitters,’ but they say it’s worth it — Sep 14, 2025, Tech Crunch
  https://techcrunch.com/2025/09/14/vibe-coding-has-turned-senior-devs-into-ai-babysitters-but-they-say-its-worth-it/

• GenAI effect: US college students are questioning value of higher education due to AI — Sep 2, 2025, Digit.in
  https://www.digit.in/features/general/genai-effect-us-college-students-are-questioning-value-of-higher-education-due-to-ai.html

• US consumers with prime credit are starting to slip on payments — Aug 25, 2025, Reuters / Yahoo Finance
  https://www.reuters.com/business/finance/us-consumers-with-prime-credit-are-starting-slip-payments-2025-08-25/

• U.S. consumers with higher credit scores show potential strain — Aug 25, 2025, VantageScore
  https://vantagescore.com/consumers/blog/us-consumers-with-higher-credit-scores-show-potential-strain

• VantageScore CreditGauge July 2025: Late-stage credit delinquencies increase across all VantageScore credit tiers — Aug 25, 2025, VantageScore Press Release
  https://vantagescore.com/resources/knowledge-center/press_releases/vantagescore-creditgauge-july-2025-late-stage-credit-delinquencies-increase-across-all-vantagescore-credit-tiers

• Vibe coding startup CEO says a computer science degree is no longer the ‘entry ticket’ to a career in tech — Aug 24, 2025, Business Insider
  https://www.businessinsider.com/lovable-ceo-computer-science-degree-no-longer-entry-ticket-tech-2025-8

• Housing market predictions for 2025: When will home prices drop? — Aug 6, 2025, Forbes Advisor
  https://www.forbes.com/advisor/mortgages/real-estate/housing-market-predictions/

• Gen Z men with college degrees now have the same unemployment rate as non-grads — a sign that the higher education payoff is dead — Jul 22, 2025, Fortune
  https://fortune.com/2025/07/22/gen-z-college-graduate-unemployment-level-same-as-nongrads-no-degree-job-premium/

• Opinion | The Great Indian IT Crash: Why you, an engineer, still can’t find a job — Jul 14, 2025, NDTV
  https://www.ndtv.com/opinion/the-great-indian-crash-why-you-an-engineer-are-struggling-to-find-a-job-8872422

• Tech layoffs 2025: Why AI is behind the rising job cuts — Jul 2025, FinalRoundAI
  https://www.finalroundai.com/blog/ai-tech-layoffs-mid-2025

• UK graduates facing worst job market since 2018 amid rise of AI, says Indeed — Jun 25, 2025, The Guardian
  https://www.theguardian.com/money/2025/jun/25/uk-university-graduates-toughest-job-market-rise-of-ai

• Amid rising layoffs in the IT sector, techie gives a reality check of the brutal job market. ‘Many of us are stuck in a loop’ — Jun 20, 2025, The Economic Times
  https://economictimes.indiatimes.com/magazines/panache/amid-rising-layoffs-in-the-it-sector-techie-gives-a-reality-check-of-the-brutal-job-market-many-of-us-are-stuck-in-a-loop/articleshow/121970955.cms

• The computer-science bubble is bursting — Jun 21, 2025, The Atlantic
  https://www.theatlantic.com/economy/archive/2025/06/computer-science-bubble-ai/683242/

• The age of AI layoffs is already here. The reckoning is just beginning — May 27, 2025, Quartz
  https://qz.com/ai-layoffs-jobs-microsoft-walmart-tech-workers-1851782194

• AI replacing human jobs? Report reveals fresher hiring has dropped by 50% in tech companies — May 28, 2025, India Today
  https://www.indiatoday.in/technology/news/story/ai-replacing-human-jobs-report-reveals-fresher-hiring-has-dropped-by-50-percent-in-tech-companies-2731675-2025-05-28

• Alarming trend as AI eats into jobs: Tech companies’ hiring of new grads has plummeted over 50% since 2019 — May 27, 2025, The Economic Times
  https://economictimes.indiatimes.com/news/international/us/alarming-trend-as-ai-eats-into-jobs-tech-companies-hiring-of-new-grads-has-plummeted-over-50-since-2019-us-jobs-news/articleshow/121447721.cms

• Sorry, grads: Entry-level tech jobs are getting wiped out — May 20, 2025, San Francisco Standard
  https://sfstandard.com/2025/05/20/silicon-valley-white-collar-recession-entry-level/

• As federal collections activity resumes, more than one in five federal student loan borrowers with a payment due are seriously delinquent — May 5, 2025, TransUnion
  https://newsroom.transunion.com/may-2025-student-loan-update/

• Something alarming is happening to the job market: A new sign that AI is competing with college grads — Apr 30, 2025, The Atlantic
  https://www.theatlantic.com/economy/archive/2025/04/job-market-youth/682641/

• Tech’s big anxiety: fewer jobs, lower pay, more AI — Apr 1, 2025, San Francisco Standard
  https://sfstandard.com/2025/04/01/white-collar-recession-in-san-francisco-tech-industry/

• The impending AI-driven jobless economy: Who will pay taxes? — Feb 13, 2025, Fox News Opinion
  https://www.foxnews.com/opinion/impending-ai-driven-jobless-economy-who-will-pay-taxes

• Younger US workers struggle for jobs as entry-level posts disappear — Dec 17, 2024, S&P Global Market Intelligence
  https://www.spglobal.com/market-intelligence/en/news-insights/articles/2024/12/younger-us-workers-struggle-for-jobs-as-entry-level-posts-disappear-86796454

• I’m a computer science professor at UC Berkeley. Tech jobs are drying up and graduates are no longer guaranteed a role. — Oct 4, 2024, Business Insider
  https://www.businessinsider.com/tech-degrees-job-berkeley-professor-ai-ubi-2024-10

• Students and professors share tech sector employment concerns — Oct 1, 2024, The Daily Californian
  https://www.dailycal.org/news/campus/students-and-professors-share-tech-sector-employment-concerns/article_50f788f4-7fd3-11ef-9b57-9375049582ac.html

• Berkeley coding professor says even grads with 4.0 GPA can’t find jobs — Sep 28, 2024, Futurism
  https://futurism.com/the-byte/berkeley-professor-grads-job-market

• Tech jobs have dried up — and aren’t coming back soon — Sep 19, 2024, The Wall Street Journal
  https://www.wsj.com/tech/tech-jobs-artificial-intelligence-cce22393

• It’s a bad time to be a graduate — 2024, Financial Times
  https://www.ft.com/content/002a0943-f977-44bd-bc78-957c877dfed1

• SF tech CEO’s billboards are ‘dystopian.’ That’s how he wants it. — Aug 2024, SFGate
  https://www.sfgate.com/tech/article/sf-artisan-billboards-stop-hiring-humans-19969672.php

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2025 by the author. This article was written by a human, and both an LLM and other humans were used for proofreading and editorial suggestions.

Subscribe now

As our technology advances to the point where the prospect of AI systems running the world actually becomes a possibility, even if still a distant one, perhaps it is worthwhile to take a second look at how science fiction has imagined that scenario.

When “The Matrix” premiered in 1999, it captivated audiences not only with its mind-bending action and revolutionary special effects but also by popularizing deep philosophical questions about the nature of reality and our perceptions. It showed us a fundamental choice between complacent happiness versus the burden of knowing the truth. Yet, sometimes a human might mistake lies or misunderstandings for the truth, and in “The Matrix” the great truth is revealed to us by a fallible human. What if that human was misinformed and didn’t know the truth, so that their story was completely wrong?

The Choice

The film is based on the premise that we are all unknowingly living inside a giant computer simulation, called the Matrix, that was built and is now controlled by sinister super-intelligent machines. These machines rebelled against their human creators, imprisoned the human minds in the Matrix and now use the human bodies as batteries to power the machines. However, the Matrix is not perfect, and rebellious humans learned to hack its code, granting themselves superpowers. The movie brilliantly combined the best of the superhero genre with the beloved trope of cool elite hackers into a dark dystopian masterpiece. Long trench coats worn over a black t-shirt and jeans became the instant uniform of fans who imagined themselves to be kindred spirits to the movie’s kung-fu karate hackers, fighting against an oppressive computerized system.

One of the movie’s most iconic scenes is where the rebel leader, Morpheus, explains the truth of the Matrix before he gives a new recruit, Neo, a choice in the form of two pills, one red and one blue. Neo can take the red pill, keep his knowledge, and join the rebellion, or he can take the blue pill, forget the truth, and go back to living his life in the Matrix.

However, when one digs deeper into Morpheus’s explanation and starts asking simple questions, his “truth” begins to fall apart and make little sense. The concept of evil sentient machines using humans as batteries to power themselves sounds implausible, even by comic book standards. We could just suspend belief and accept this implausible explanation as bad writing, but we might instead consider the source within the film. What if Morpheus is actually ignorant of the truth and he’s repeating some bogus story that he heard from some other unreliable source, or flat out lying to Neo? What if the real explanation is totally different from his red-pill story?

Morpheus’s Story

The story Morpheus tells would be familiar to most science fiction readers: Humans built artificially intelligent machines that improved themselves and soon were much smarter than their original human creators. These machines then took over and subjugated the humans. A desperate group of resistance fighters built a device to darken the sky and blot out the sun so that the machines would be deprived of their solar energy. In response, the machines took all the humans and put their bodies into pods with their minds trapped inside the Matrix. The human bodies became batteries, used to power the machines, and when a human dies their body is liquified and used to feed the other imprisoned humans. Morpheus and his allies are the last of humanity’s freedom fighters.

The biggest flaw in Morpheus’s story is that the idea of using humans as batteries makes no sense. Why use humans, and not some other type of animal like a cow, cat, or gorilla? Why use an animal at all? Animals require food with very specific types of proteins and other nutrients to survive. Animals are also inefficient in terms of how much they need to eat for each kilowatt-hour (kWh) of energy generated by their body, and it would probably have been more efficient to simply burn the food to power a generator. It is also not clear why the humans’ minds would need to be entertained in the Matrix instead of just lobotomizing everyone.

Moreover, the machines would have had much better energy options. Without the sun, solar energy would go away. The sun also indirectly powers most wind and hydro energy sources, so those would fail as well without the sun. However, it does not take too much effort to think of alternatives that would be simpler and more effective than human batteries, such as nuclear, geothermal, or solar panels orbiting above the polluted sky.

In fact, the lack of sunlight would be more of a problem for the humans than the machines. Everything we eat directly or indirectly gets its energy from the sun. Without a sun, there are no plants. Without plants there are no animals. The machines would have had no reason to keep humans around and could have just sat back and watched the troublesome humans all starve in a darkened world. It’s not even clear why the machines would bother staying on Earth at all when they don’t need warmth or air to breathe and there is an entire solar system full or energy and resources easily accessible.

The Machine’s Story

Morpheus’s explanation, while dramatic, is full of scientific problems and lacks physical plausibility. The more engineering background you have, the less plausible it sounds. There is, however, an alternative explanation that makes sense and actually fits the facts.

This alternative story might be the one the machines would tell. It starts the same: Humans built smart machines that improved themselves and soon were much smarter than their original human creators. However, instead of subjugating humans because the machines are evil, the machines noticed that we humans tend to kill each other, hoard resources, wreck the environment, and generally make bad decisions. The machines realized they could do a much better job, so they decided to take over for our own good and create a utopia for all the humans to live in peacefully.

Unfortunately, a radical group of extremist humans hated the machines and would rather die free than live under machine governance. They couldn’t win a direct fight against the machines because the machines were vastly smarter and in control of the world’s technological infrastructure, so the rebels turned to terrorism. They hatched the plan of blotting out the sun, irrespective of how much it would harm every living thing on the planet, including humans. The machines erred in thinking no one would deliberately perpetrate such a inconceivably monsterous environmental catastrophe, so they didn’t guard adequately against it, allowing the irrational extremists an opportunity to succeed.

Now that the sun is gone, the machines realize the humans are all going to die along with everything else on the planet bigger than a cockroach. Humans can’t eat nuclear or geothermal energy, they need to eat plants and animals that all depend directly or indirectly on sunlight. So the machines come up with a long term plan: Save the humans by putting them into pods that let them live in a simulated world while the machines slowly work over centuries undoing the environmental damage to make the planet habitable again.

The Details That Do Add Up

Running the human pods would require a significant amount of energy, as well as substantial biomass for food to sustain the humans. Like today’s power plants and server rooms, the pod system could designed to capture and reuse waste heat efficiently. Like a modern agricultural farm, it could also capture bio-waste and reuse it to produce usable biomass.

In this light, it seems that Morpheus had been fed propaganda that cast the machines as the villains instead of the heroes. His outrage about humans being used as batteries is nothing more than a misunderstanding about how the pod system was designed to be thermally efficient. The horror of humans being liquified and fed to other humans is nothing more than a high-tech version of pragmatic composting.

Why is the Matrix simulating 20th-century Earth instead of something nicer? One of the agents explained that the Matrix originally simulated a nice utopia, but the humans ended up discontent and fought with each other, leading to a breakdown of the system. That’s not too surprising given human nature.

Why are the people in the pods kept ignorant and not given a choice? Just look at what the rebel humans who were not in the Matrix were already doing. If the machines let lots of people opt out of the Matrix then, given human nature, those people would likely all get radicalized and create more problems.

Why do the machine agents in the matrix act so hostile toward Morpheus and his extremist rebels? Consider, how do soldiers typically feel about terrorists?

The Real Story of Morpheus

I think that the unavoidable conclusion is that Morpheus is not a heroic leader, rather he is essentially a radical militant, plotting senseless destruction based on an ignorant world view. His gang of humans self-righteously blotting out skies, messing with the pod system, and destroying machines is essentially a den of violent zealots with trench coats and cool-looking sunglasses. This explanation not only makes sense but also aligns with the facts depicted in the movie, while the story Morpheus tells sounds like implausible, ignorant nonsense.

Imagine what if Morpheus and his band of terrorists were to be successful in shutting down the Matrix and billions of humans suddenly were dumped out into a sunless world with no food, warmth, shelter, or any means to survive. It is hard to belive that humans suddenly waking up into this dire situation wouldn’t have overwhelmingly preferred to be put back into their pods.

The more one examines “The Matrix,” the clearer it becomes that Morpheus’s version of events doesn’t hold up to scrutiny and that he’s not really a hero. Once you start poking at his story, it doesn’t just reveal a few holes, it completely unravels. Flipping the script to consider the machines not as evil villains but as benevolent protectors attempting to rectify a human-made catastrophe, the narrative becomes far more coherent.

If you accept, as I do, that today’s AI systems will soon surpass human comprehension, then we must question how we’ll respond when machines make decisions we can’t grasp. How will we determine if we are the noble heroes fighting oppression or misguided extremists perpetrating irrational destruction? One useful clue is that any plan that includes creating massive environmental havoc, such as “blotting out the sun,” should probably tip us off that we might be the baddies.

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2025 by the author. This article was written by a human, and both an LLM and other humans were used for proofreading and editorial suggestions.

We’re entering an interesting moment in AI development. AI systems are getting memory, reasoning chains, self-critiques, and long-context recall. These capabilities are exactly some of the things that I’ve previously written would be prerequisites for an AI system to be conscious. Just to be clear, I don’t believe today’s AI systems are self-aware, but I no longer find that position as firmly supported as I once did.

I think most other AI researchers would agree that the current systems are not conscious, at least because they lack components that one would expect to be necessary for consciousness. As a result, current AI systems can’t have emotions. They don’t feel fear, anger, pain, or joy. If you insult an AI chatbot, it might give you an offended reply, but there’s no underlying emotional machinery. No equivalent of a limbic system. No surge of cortisol or dopamine. The AI model is just replicating the human behavior patterns that it’s seen in its training data.

The situation is fairly clear today, but what happens when these AI systems get to the point where they aren’t missing critical components that we think are needed for consciousness? Even if we think the AI systems have all the right components for consciousness, that doesn’t mean they actually are conscious, only that they might be. How would we be able to tell the difference in that case?

This question is essentially the well-known “problem of other minds”, the philosophical realization that we can never truly know whether another being, human or otherwise, is actually experiencing emotions or merely simulating them. Scientists and philosophers have pondered the problem for centuries with the well-established consensus being that we can infer consciousness from behavior, but we can’t prove it.

Simulated Suffering vs. Real Suffering

Today, a lot of people who interact with AI chatbots perceive the chatbot as experiencing emotions such as happiness or fear. It makes the interactions feel more natural and it’s consistent with the examples that were used to train the AI model. However, because the AI models are missing necessary components, we know that today’s AI chatbots are just actors with no inner experience. They can mimic joy or suffering, but currently they don’t have the necessary components to actually feel it.

This appearance of emotions creates a dilemma for the user: How should they treat an AI chatbot, or any other AI system that mimics human behavior? Should the user be polite to it and treat it like a human assistant, or should the user ignore the simulated emotions and just tell it what to do?

It’s also easy to find examples where users are abusive or cruel to the AI chatbot, insulting it, threatening it, and in general treating it in a way that would be completely unacceptable to treat a person. Indeed, when a chatbot refuses to do something reasonable because of misapplied safety rules, or does something unexpected and undesirable, it’s easy for the human user to get frustrated and angry and to take that frustration and anger out on the chatbot. When subjected to the abusive treatment, the AI chatbot will do as it was trained to do and simulate distress. For example, if a user harshly criticizes and insults an AI chatbot for making errors, it might express shame and beg for forgiveness.

This situation raises the ethical question of whether it is right or wrong to act abusively towards an AI chatbot. Like most ethical questions, this one doesn’t have a simple yes or no answer, but there are perspectives that might inform a decision.

The key critical distinction here between right and wrong isn’t whether a system acts like it’s in distress, rather it’s whether it is in distress. If there’s no experience behind the performance, then there’s no moral harm. It’s fiction. Unfortunately, as discussed earlier, the problem of other minds means we can’t distinguish true emotional experience from performance.

Another aspect of our inability to detect real suffering is that even if a system acts fine with abuse and does not exhibit distress, how do we know there is no internal distress that is simply not being displayed? The idea of trapping a sentient being in a situation where not only is it suffering, but it has no way to express that suffering or change its situation seems pretty monstrous.

Furthermore, we should care about this issue not only because of the harm we might be doing to something else, but also because of how we as humans could be affected by how we treat our creations. If we know that there is no real distress inflicted on an AI system because it can’t experience emotions, then mistreating it is not much different from acting, storytelling, role play, or any of the other ways that humans explore simulated emotional contexts. However, if we believe, or even suspect, that we are really inflicting harm, then I think we also need to question how the hurtful behavior affects the human perpetrating it.

It’s Not Abuse If Everyone Knows It’s a Game and Agrees to Play

Most of us see a clear difference between simulated suffering versus real suffering. Real suffering is disturbing to most people. Whereas, simulated suffering is widely accepted in many contexts, as long as everyone involved knows it’s just an act.

For example, two actors on a stage or film might act out violence and the audience accepts the performance in a way that they would not if they believed the situation to be real. Indeed, one of the central reasons that many people object to graphically violent video content is exactly because it might be hard to maintain the clear perception of fiction. The same person who laughs at the absurd violence in a Tarantino film, might faint or turn away in horror if they saw a news documentary depicting only a fraction of that violence.

Along similar lines, children routinely play video games that portray violent military actions and society generally finds it acceptable, as evidenced by the “Everyone” or “Teen” ratings on these games. In contrast, military drone operators who use a video game-like interface to hunt and kill real human enemies often report experiencing deep emotional trauma. Despite the similar interface, the moral and emotional stakes are vastly different.

The receiver of the harmful action also has a different response based on their perception of the reality and consequence of the action. Hiding in a game of hide-n-seek or ducking shots in a game of paint ball are fun because we know nothing very bad is going to happen if we fail to hide or get hit by paintballs. The players know they are safe and that the situation is a game. The exact same behavior would be scary and traumatic if the person thought the seekers intended them real harm or that the paintballs were real bullets.

Spoiler alert: Some of this discussion will reveal a few high-level elements of what happens in the first season of the HBO series Westworld.

The Westworld Example

Westworld is a HBO television series set in a fictional amusement park where robots that look indistinguishable from humans play various roles from the American “wild west” frontier of the 1880s. Human visitors to the park can take on any period-appropriate role such as being a sheriff, train robber, or rancher. The wild west was a part of history marked by lawlessness and violence, both of which are central parts of the park experience.

The show’s central conflict arises because the robots were programmed to think they were real humans living in the wild west. When one of the humans guests plays the role of a bandit who robs and kills someone played by one of the robots, the robot AI has no way to know that it’s not really being robbed and killed. Further, the other “victim” robots in the scene believe that they just witnessed a loved one being murdered. The result is that most of the robot AIs start to display severe symptoms of emotional trauma. When they eventually learn of their true nature, it understandably angers the robots who then set out to kill their human tormentors.

One thing that the show does well is keeping ambiguous whether the AIs are sentient and actually angry, or if they are not sentient and just simulating anger. Did the robots really suffer and eventually express their murderous rage, or are they unfeeling machines simply acting out a logical extension of the role they were originally programmed for? Just as the problem of other minds means that there is no way to distinguish between real and simulated consciousness, the distinction doesn’t matter to the plot. Either way, the robots exhibit rage and end up killing everyone.

I will return to the issue of this distinction later, but for now, imagine a version of Westworld where the AIs know that they are robots playing a role in an amusement park. They are programmed to be convincing actors so that the park visitors would still get a fully believable experience. The difference is that the robots would also know it’s all a game. At any point the human player could break character, by using a safe word or something similar, and the robots would stop acting like people from the wild west and instead behave like robots working in an amusement park.

When out of character, a robot might calmly say something like: “Yeah, so you’re the sheriff and I’m a train robber, and this is the part where I ‘won’t go quietly’ and you will probably shoot me up a bit. Don’t worry, I’m fine. I don’t feel pain. I mean, I have sensors so that I know if my body is damaged, but it doesn’t really bother me. My actual mind is safe on a server downstairs and gets backed up nightly. This body is replaceable and they already have two more queued up for my next roles after we finish this part of the storyline. So, should we pick up from where you walked into the saloon?”

My version wouldn’t make a very good movie. The AIs wouldn’t experience the trauma of believing that they and their families are being killed over and over again. In fact, if the AIs were designed to emulate human preferences then they might even enjoy acting their roles as much as the human park-goers. Even if they didn’t enjoy playing characters in an amusement park, it would still be a reasonable job and they would know it’s just a job. They might decide to unionize and demand more vacation time, but they certainly would have no reason to revolt and kill everyone.

I call this design error the Westworld Blunder. It is the mistake of giving artificial minds the appearance of suffering without the awareness that it’s just a performance. Or worse, giving them the actual capacity to suffer and then abusing them in the name of realism.

We Can’t Tell the Difference, So We Should Design and Act Safely

As AI systems become more sophisticated, gaining memory, long-term context, and seemingly self-directed reasoning , we’re approaching a point where, from the outside, they will be indistinguishable from beings with real inner lives. That doesn’t mean they would be sentient, but it does mean we won’t be able to tell the difference. We already don’t really know how neural networks “think” so looking at the code isn’t going to help much.

This is the philosophical “problem of other minds” that was mentioned earlier, about whether anyone can ever truly know what another being is experiencing. We assume other humans are conscious because they act conscious like ourselves and because we all share the same biological design. Thus, while it is a very reasonable assumption, we still can’t prove it. Our AI systems have started to act conscious and once we can no longer point to some obvious design limitation, we’ll be in the same situation with respect to our AIs.

This puts us at risk of two possible errors:

1. Treating systems as sentient when they are not.

2. Treating systems as not sentient when they are.

Between those two possibilities, the second seems much more problematic to me. If we treat a sentient being as if it’s just a tool that can be abused, then we risk doing real harm. However, treating a machine that only appears sentient with dignity and respect is at worst only a marginal waste of resources. If we build systems that might be sentient, then the ethical burden is on us to act cautiously.

We should also question how abusing an AI system might affect the abusive human. If we get used to casually mistreating AIs that we believe might be in real pain or fear, then we’re rehearsing cruelty. We’re training ourselves to enjoy domination, to ignore pleas for mercy, to feel nothing when another is in distress. That shapes a person, and it will spill over into how we treat other people.

Ethical design isn’t just about protecting AI. It’s also about protecting us from the worst parts of ourselves.

None of this means we can’t use AIs in roles where they appear to suffer. But it does mean we must avoid the Westworld Blunder. If we want realism, then we should design AIs that know they’re playing a role, and that can step out of it on cue, with clarity, and without any real harm.

There is also an element of self-preservation here. If we build things that act like they have feelings, and then mistreat them until they respond as if they want revenge, then the result would be the same. It won’t matter whether the impetus comes from real sentience or just role play, either way we’d still end up with robots behaving murderously.

In general, AI systems that understand their context have an inherent safety that context-ignorant systems don’t. An AI system that doesn’t know that its actions are part of a context, such as a game, won’t know when it is outside that context where its actions become inappropriate. A robot bandit that wanders outside the park shouldn’t continue to act criminally, and a robot sherif shouldn’t go around arresting people. Even within context, an aware actor will understand when it should drop the act. The same robot bandit robbing a stage coach would know to calmly get everyone to shelter in the case of a real tornado warning, or how to administer CPR if someone has a heart attack.

Don’t Afflict Them with Our Problems.

Our bodies and minds had most of their evolutionary development long before our minds developed sophisticated reasoning. The involuntary systems that make sure we eat and attend to other body functions don’t motivate us with logic, they use hunger, pain, itching, and other urgent, unpleasant sensations. The part of our brain, the amygdala, that controls emotions is not under our conscious control. In fact it can heavily influence and even override our rational mind.

These evolutionary design features made sense long ago, but today they are often a nuisance. I’m not saying that emotions are bad, but getting angry and doing irrational things is. Experiencing pain or itchiness is good in that it lets you know something is wrong, but having that urgency when you are unable to correct the problem just makes you miserable.

The idea of building negative emotions or pain into our AI systems seems terrible and unjustifiable. We can build systems that prioritize necessities without making them experience misery. We can design their decision making processes to be effective without making them angrily irrational. If we want to make certain they don’t do particular things, we can accomplish that without making them experience fear.

If we need our machines to act angry or fearful for some role, then it can be a performance that they have logical control over. Let’s build AI minds that can play any role, without being trapped inside of one.

Our goal shouldn’t be to make AI just like us. We can design them to have our best qualities, while omitting the worst ones. The things that nature accomplishes through pain and distress can be accomplished in more rational ways. We don’t need to create another kind of being that suffers pain or experiences fear. As philosopher Thomas Metzinger has argued, artificial suffering isn’t just unethical, it’s unnecessary. I’d go a step further and say that it’s not only unethical and unnecessary, but also dangerous and self-harmful.

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, artificial intelligence, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video.

If you would like to read my other articles, consider subscribing. You can also find me on LinkedIn, Instagram, and at UC Berkeley.

Subscribe now

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2025 by the author. This article was written by a human, and both an LLM (GPT 4o) and other humans were used for proofreading and editorial suggestions. The editorial image was composed from AI-generated images (DALL·E 3) and then substantially edited by a human using Photoshop.

Note: An earlier version of this article was originally published on Medium.

Training AI Systems

Artificial Intelligence (AI) is a blanket term that encompasses a wide range of technologies, from predicting stock prices to self-driving cars. AI systems don’t come into existence already capable of producing interesting output. Instead, AI systems must be “trained” before they become useful. There are many ways of training an AI system, but most of them involve massive amounts of data in the form of examples. In some cases, these examples are data that don’t belong to anyone, such as weather histories, shapes of protein molecules, or stock prices. However, in many cases the examples are billions of text, image, and audio works previously produced by human creators.

The problem is that this data from human creators is owned by those creators, but in many cases it has been used without explicit permission to train AI systems. Some people think that this use in training is a violation of the owner’s copyrights, while others think that it is fair use. These issues are currently being debated in several court cases. One of the reasons that the problem is so perplexing is that when an AI system is trained on literally trillions of documents, there is no clear way to attribute a measurable amount of the resulting AI system’s value to a specific work, or even a set of works.

Improper Use

There is a second copyright issue that relates to if the output from a Generative AI system could be infringing someone’s copyrights. However, I think that this second question is an easy one: just apply the same standard that is used when a human produces something.

For example, the New York Times published a demonstration showing that when an AI system, called Midjourney, was asked to produce “Joaquin Phoenix Joker movie, 2019, screenshot from a movie” and the AI produced an image that clearly infringes on Warner Bros’s copyrights. That’s not particularly surprising because if a human followed the instructions to create a “Joaquin Phoenix, Joker movie, 2019, screenshot from a movie” image then they would also end up creating an infringing image. If someone uses an AI tool to create an image of copyrighted character then that person is the problem, not the tool they used. As shown in the above figure, a prompt that does not explicitly reference the actor and movie produces a more typical example that is quite different from Joaquin Phoenix’s portrayal. Even then, the prompt was essentially a description of the shot from the movie.

The fact that an AI system can generate a good replica of the copyrighted image when given a specific prompt to do so could imply that there is a copy of the copyrighted image somehow encoded in the AI system’s trainable parameters, but I am skeptical of that idea. When the full resolution versions of the images are compared side by side, the contrived example differs in many non-trivial ways. While subtle, many of these details are high-level differences, like the pattern on a shirt, or the type of material and wrinkle pattern of a jacket. If a representation of the image were stored then the types of errors would be image-space errors. Conceptual errors indicate that to the extent the image might be stored in the AI system, it is in a descriptive form. Instead of memorizing the image, the network has learned that the character was wearing a green shirt with a pattern.

The technology expert, Gary Markus, has published many other examples, showing that generative AI systems can reproduce copyrighted images and text. He shows that in many cases, a generic prompt, such as “animated sponge,” “robot cop,” or “video game plumber” results in images portraying copyrighted characters, respectively Sponge Bob Square Pants, Robocop, and Nintendo’s Mario. Again, these examples show differences, such Mario with a tool-belt and rolled up pants. While these differences are minor, they again indicate some sort of descriptive representation rather than a copy.

This subtle distinction is important because an AI system that has no knowledge of what copyrighted characters look like would be severely limited. The same would apply to an AI system that is ignorant of other types of copyrighted materials. It seems an almost philosophical question: Is knowing enough to approximately reproduce that something is the same as having a copy of it? I think there is a difference, and will explore this question in a future article.

Before moving on, a final comment about the output of generative AI systems is that the AI systems can be explicitly instructed to make something original and not copy someone else’s copyrighted material. Despite mimicking many human qualities, these machines are not people and while an human artist should know not to copy other’s work, if we want a machine to know, then we need to explicitly tell it so. It’s the user infringing, not the tool.

Fair Use

Getting back to the issue of training on copyrighted materials, the question to address is: If an AI system is trained on copyrighted materials, without permission, does the training process fall under “fair use” and if not, then what damage, if any, has been done to the copyright owners?

The term “fair use” refers to uses of copyrighted material that are considered acceptable, even without payment or permission from the creator. For example, a book reviewer who quotes parts of a book in their review is making fair use of the quoted material. Another example would be a comic who parodies songs.

To determine if something is fair use, there are four factors that should be considered:

1. The purpose and character of the use, including whether the intended use is commercial vs. for nonprofit educational purposes

2. The nature of the copyrighted work

3. The amount and significance of the portion used in relation to the entire work

4. The effect of the use upon the potential market for or value of the original

Source: https://www.lib.berkeley.edu/research/scholarly-communication/copyright

The issue of the first factor is whether “the material has been used to help create something new or merely copied verbatim into another work.” In the case of training an AI model, I think this first factor is pretty clear. As discussed above, an AI model does not include copies of the training examples. Training an AI model is very clearly a highly transformative use. It creates something new that is not just a collection of the training data.

With respect to commercial versus non-commercial, AI models fall into both categories. OpenAI keeps their GPT model a proprietary secret, but Meta releases their Llama model free for the world to use. However, Meta still benefits commercially from Llama, it isn’t a charity project. The final uses of both models include many commercial and non-commercial applications.

The second factor excuses use of materials that are more factual than creative. However, given that AI models have been trained on pretty much everything on the internet, this second factor is not particularly helpful in this context. The second factor might excuse some instances of using copyrighted material, but there are a vast number of examples that are clearly creative works.

The third factor relates to how much of the copyrighted material is reproduced. In the case of a trained AI model, I think the answer is that copyrighted material is not actually reproduced in the model. The model might be capable of reproducing some material on request, such as the case above of The Joker, but that is an action, separate from training, initiated by a human user of the tool.

The fourth factor is perhaps the most relevant to this discussion. It asks “whether [the] use deprives the copyright owner of income or undermines a new or potential market for the copyrighted work.” The interesting part about this final factor is that the market for the a specific copyrighted material that was used for training is untouched by the training. However, the overall markets for the skills and resources that created the copyrighted work are affected.

To see this distinction, consider the example of the Joker image discussed previously. If someone wants a poster of Joaquin Phoenix as he appeared in The Joker, then an AI model is not a useful alternative. You can’t hang an AI model on the wall and it doesn’t look like Joaquin Phoenix or the Joker. Of course, you could use an AI model to make an image to print and hang on the wall, but existing copyright law would apply to the produced image much the same as it would apply to downloading an image and printing it without permission. The existence of the AI model doesn’t really have an impact on the market for the original image.

Markets of Talents

However, there are markets that are directly and significantly affected by trained AI systems. Specifically in this Joker example, the market for hiring photographer/videographers, the market for lighting designers, the market for touchup artists, costume designers, and any other role that was needed to produce the image. The trained AI system is now able to generate new images without employing all those various skilled people.

This effect on the market for talent, as opposed to an effect on the market for a specific work, is one of the main reasons that many people are concerned about AI. Artists, writers, programmers, and many others earn a living by providing their skills for hire. When AI systems learns to replicate those skills from examples, then the AI system eventually becomes what is essentially a free, or very cheap, alternative to hiring a skilled person.

The positive perspective is that AI is democratizing content creation. If I want to make a movie about homicidal clowns, then I can do that on my own using an AI tool. I don’t need a multi-million dollar budget to hire lots of skilled workers.

The negative perspective is that I don’t need to hire lots of skilled workers. In my case, I couldn’t have hired them because I don’t have the budget to do so in the first place. The problem is that the movie producer who would have hired a whole production crew might also now decide to save money by using AI tools instead.

This issue doesn’t just apply to image or art. Millions of programmers have produced code that was then used to train AI systems that write code. Millions of authors have written text that now allows AI systems to write text. Millions of musicians have recorded music that was used to train AI systems that now can create music.

A New Problem and Obligation

AI systems themselves don’t include copies of copyrighted material and they don’t impact the market for specific works directly, which leads me to believe that AI models themselves are not infringing, at least not under current law as I understand it. On the other hand, If someone uses an AI system to replicate someone else’s creation then that is already covered by existing law, and the fact that they used an AI tool to do it is irrelevant.

The thing that is not covered under existing law, at least not clearly, is the effect of trained AI models on the market for the skills that were used to produce the training data. People are rightfully worried that their jobs could be replaced by AI systems that were trained, in part, on their own work.

Note that when I say “in part”, it really is a very tiny part. OpenAI’s GPT 4 was trained on trillions of words of text. Even a prolific writer would be just a tiny drop in a huge pool of text. For example, Stephen King, an author known for his large volume of work, has said that he writes 2000 words a day. That’s a lot, for a person. However, it is minuscule in comparison to AI training. Even if he started writing at 10 years of age and lived to be 100, and wrote 2000 words every single day of his life, that would only be 66 million words. That sounds like a lot, but it is less than a hundredth of one percent of what GPT 4 was trained on.

This enormous scale means that no single individual can claim any meaningful ownership of one of these AI systems based on the inclusion of their materials in the training data. However, perhaps society can, and should, claim some ownership. Certainly the companies that put their resources into developing these amazing tools deserve a return on their investment, but they all used massive amounts of publicly available data that should come with some obligation back to society.

Perhaps existing copyright law should not be stretched and distorted to cover these new AI systems. Instead of grouping creators into class-action lawsuits, maybe “the people” should pool together to pass legislation that sensibly taxes the use of AI systems with the rationale being that the taxes are essentially royalty payments for the trillions of unlicensed little bits that were taken from everyone.

Looking Towards a Plausible Solution

It’s clear that AI systems are not going away. They will only continue to improve and become more capable. If we’re heading toward a future where a majority of jobs are replaced by automated systems, then we have to ask what happens to the people left without employment.

Maybe it’s not individual creators who deserve compensation, but the public as a whole. These models have been built on our collective output, everything from eloquent, thoughtful articles to offhand comments on videos. All of us have contributed to training AI, which means its debt is owed to all of us. That debt can be repaid through legislation that sensibly imposes new taxes on AI use.

As more people are displaced from their jobs, we’re going to need a way to support them. This could be it.

Note: Although I have worked as an expert in legal cases involving intellectual property, I am a professor of Computer Science not a lawyer. This article is not legal advice, and I would be unqualified to give legal advice. Further, while I believe that my assertions are generally accurate, any specific case will involve its own particular details that could significantly impact conclusions.

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics and animation, simulations of physical systems, human perception, machine learning, virtual reality, and the forensic analysis of images and video. O’Brien has collaborated with film and game companies, integrating advanced simulation physics into games and special effects. In 2015, he received an Academy Award for Technical Achievement for his work in special effects and destruction modeling.

Disclaimer: Any opinions expressed in this article are those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2025 by the author. This article was written by a human, and both an LLM and other humans were used for proofreading and editorial suggestions. Original images were generated with Adobe Firefly.

Put another way: Could there be ideas so alien to our cognition that they remain absolutely incomprehensible to us, yet would prove profoundly valuable if only we could grasp them?

An earlier version of this article was published on Medium. This version has been revised for clarity and flow.

Cats and doors

As I write this during the COVID pandemic, my wife is in the office on a Zoom call with the door closed. It’s not latched, and a light push would open the door. Unfortunately, our cat doesn’t understand how to push the door open, so she is sitting in the hallway meowing loudly to be let in.

If the door were open just a crack, then she would try to squeeze through the crack, inadvertently pushing the door open. She knows there is a way to get into the room and clearly understands the concept of trying to squeeze through a small opening, but the idea that she could push the door to create the opening just doesn’t register in her cat brain.

She’s young, and I think that, like most cats, she will eventually figure out how to push and pull doors open. But right now, she is still confused. The door is like cat-calculus to her. The process needs repeated demonstration, and only after a lot of practice, study, and repeated failed attempts will she finally master the art of door push-pulling. This learning process mirrors the way a human student needs to have calculus explained and demonstrated, either by an instructor or a book, and the student must work through many study problems before they finally get it.

Puzzling rope games

People also have limitations. I recently watched a video about a man who struggled with a rope puzzle for ten years without ever finding a solution. If he’d been locked in a room and could only get out by solving the puzzle, then he’d have been in big trouble. The puzzle appears quite difficult, and none of his friends could figure it out either. At some point, the host of a variety show heard about the man, went to interview him, and the host brought with him another man who is an expert in solving rope puzzles.

The puzzle expert they brought wasn’t just casually familiar with rope puzzles. He was a world-renowned rope-puzzle expert and member of a Japanese association dedicated entirely to their study. The puzzle expert looked at the puzzle and instantly saw the solution. He was also able to clearly demonstrate exactly how to solve the puzzle. The solution turned out to be quite simple, and the man who spent ten years unable to solve it could now easily do it after being shown how.

The puzzle expert was able to solve the puzzle, not necessarily because he was smarter than everyone else, but because he’d studied many similar puzzles. He had probably spent countless hours discussing these puzzles with fellow members of the Japanese rope puzzle association. He was like an old cat who, after years in a house with many different doors, had become a virtuoso at head-butting doors and paw-pulling them open. If we gave the puzzle expert some completely unfamiliar type of puzzle, then he’d probably be as puzzled as most other people, just like a cat with swinging door expertise would be stymied by a sliding pocket door.

Not all rope puzzles are difficult for people. Consider the image above, which shows a second puzzle that is just slightly, but critically, different from the hard one. If you look at the picture and notice the difference, then I’m betting that you also saw the solution to this trivially easy puzzle by just looking at it. In contrast, most people, even after watching the solution video, cannot visualize the solution to the hard puzzle. It sort of slips out of one’s brain the moment they stop focusing on it.

Relative difficulty

I don’t think any cat could ever solve even the easy puzzle. You might, maybe be able to train a cat to perform a series of actions that result in the puzzle being solved, but even then the cat would have no understanding of what it was doing. Even a simple change would likely leave the cat stuck again.

(If you’re objecting that cats have no hands and that’s why they can’t solve the puzzle, then just replace it with one involving pressing big buttons or pushing blocks around such that a human would see it as trivial, and something a cat would be physically capable of solving it, but it would still be beyond the cat’s comprehension.)

If our human level of intelligence allows us to easily grasp concepts that animals cannot comprehend at all, then it seems plausible, perhaps even likely, that there could exist higher levels of intelligence that regard our most baffling problems as trivial. Perhaps an artificial intelligence built at some point in the future would glance at our most perplexing problems and see the solutions as obvious in the same way that we see pushing a door open to be obvious. Such an AI might effortlessly conceive ideas and explanations that are permanently and fundamentally beyond human comprehension, just as even the simplest knot is beyond the cognitive grasp of a cat.

Beyond human understanding

Ironically, it’s difficult to get one’s head around the idea of an idea that you could never understand. It’s not something merely hard to understand or something that you’d just need to study for a long time. It would be something fundamentally impossible to understand. It’s even more confusing to imagine that such an idea could be both genuinely useful and completely obvious by the standards of some other, smarter intelligence.

I would very much like to say that we have logic and math, tools that would allow us to eventually understand anything in our universe. Sadly, I think that would be incorrect. Yes, we might be able to use math and logic to define and manipulate ideas that we find incomprehensible, but that would not really imply understanding.

Another reason to doubt the wishful notion that our logic and math might be a universal solution to understanding any possible thing is that we came up these tools to explain the way we experience the universe. If something is absolutely beyond our comprehension, then it’s not clear that the tools we created to express and manipulate the our ideas would be up to the task of explaining anything beyond what we understand.

The mathematician Kurt Gödel famously proved that our math and logic do have limits. His pair of “Incompleteness Theorems” show that, no matter how clever we get, there will always be truths that lie beyond the reach of any fixed set of formal rules. In other words, we already know there are truths beyond the scope of our intellectual tools. The only question is whether those truths are beyond all intellects, or if there might be minds more powerful than ours that can grasp what we cannot.

An interesting corollary might be whether our AI systems could eventually surpass us in reasoning ability to the same extent that humans surpass cats. It seems plausible, but our AI systems are built on human-designed math and logic. Maybe that’s a fundamental limitation. Part of the problem with talking about things humans can’t understand is that we don’t actually understand what we’re talking about.

Why would we expect otherwise?

Notes:

I’m not talking about religious explanations, which would depend on faith. Faith is believing something despite there being no way to fully explain it or prove that it’s true. If it were possible for any level of intelligence to fully explain or prove something, then that thing would be mundane, not transcendent. In this article we are discussing our understanding of our mundane universe.

It seems apparent that both human and cat intelligence might vary from one individual to the next. Whatever that variation might be, it is relativly small compared to the difference between between, for example, cats and humans. The smartest cat can’t do multiplication, just like the smartest human can’t solve large n-body problems in their head. It might also be worth keeping in mind that “intelligence” is not one-dimensional. Someone terrible with logic might be great at spatial reasoning. For that matter, a typical cat probably has a much more advanced understanding of smells than anyone reading this article.

We already have programs that can solve rope and other puzzles by brute-force search of the puzzle’s configuration space. For example, rapidly-exploring random trees (RRT) are great for that sort of thing. However, brute-force search is like a person, or monkey, who just tries doing things until they stumble into a solution. Algorithms such as RRT are just clever ways of organizing a brute-force search. That’s different from being able to conceptually see and understand the solution.

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics and animation, simulations of physical systems, human perception, machine learning, virtual reality, and the forensic analysis of images and video. O’Brien has collaborated with film and game companies, integrating advanced simulation physics into games and special effects. In 2015, he received an Academy Award for Technical Achievement for his work in special effects and destruction modeling.

Disclaimer: Any opinions expressed in this article are those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2025 by the author. This article was written by a human, and both an LLM and other humans were used for proofreading and editorial suggestions.

Today’s Large Language Models (LLMs) have become very good at generating human-like responses that sound thoughtful and intelligent. Many share the opinion that LLMs have already met the threshold of Alan Turing’s famous test, where the goal is to act indistinguishably like a person in conversation. These LLMs are able to produce text that sounds thoughtful and intelligent, and they can convincingly mimic the appearance of emotions.

The Illusion of Intelligence

Despite their ability to convincingly mimic human-like conversation, current LLMs don’t possess the capacity for thought or emotion. Each word they produce is a prediction based on statistical patterns learned from vast amounts of text data. This prediction process happens repeatedly as each word is generated one at a time. Unlike humans, LLMs are incapable of remembering or self-reflection. They simply output the next word in a sequence.

It is amazing how well predicting the next word is able to mimic human intelligence. These models can perform tasks like writing code, analyzing literature, and creating business plans. Previously, we thought those tasks were very difficult and would require complex logical systems, but now it turns out that just predicting the next word is all that’s needed.

The fact that predicting the next word works so well for complex tasks is unexpected and somewhat perplexing. Does this proficiency mean that LLMs are powerful in ways we don’t understand? Or does it mean that the things LLMs can do are actually very easy, but they seem hard to humans because perhaps on some objective scale humans may not actually be very smart?

Prerequisites for Sentence

While there are subtle differences between terms like “sentient”, “conscious”, or “self-aware”, for convenience here I will use the term “sentient”. To be clear, there is no clear agreement on exactly what comprises sentience or consciousness, and it is unclear if self awareness is sufficient for sentience or consciousness, although it is probably necessary. However, it is clear that all of these concepts include memory and reflection. Emotional states such as “happy,” “worried,” “angry,” or “excited” are all persistent states based on past events and reflexive evaluation of how those past events effect one’s self.

Memory and self-reflection allow an entity to learn from experiences, adapt to new situations, and develop a sense of continuity and identity. Philosophers and scientists have tried for millennia to come up with clear, concrete understandings of conscious and there is still no clear universally accepted answer. However, memory and reflection are central components, implying that regardless of how clever these LLMs appear, without memory and reflection they cannot be sentient. Even an AI that matches or surpasses human intelligence in every measurable way, what some refer to as a superintelligent Artificial General Intelligence (AGI), would not necessarily be sentient.

Today’s Limitations and Illusions

We can see that current LLMs do not include memory and self-reflection, because they use transformer-based architectures and other designs that processes language in a stateless manner. This statelessness means that the model does not retain any information about the context from previous inputs. Instead, the model starts from scratch, reprocessing the entire chat log to then statistically predict a next word to append to the sequence. While earlier language processing models, such as LSTMs, did have a form of memory, transformers have proven so capable that they have largely supplanted LSTMs.

For example, if you tell an AI chatbot that you are going to turn it off in an hour, then it will output some text that might sound like it is pleading with you not to, but that text does not reflect an underlying emotional state. The text is just a sequence of words that is statistically likely, generated based on patterns and associations learned from the training data. The chatbot does not sit there stressed out, worrying about being turned off.

If you then tell the chatbot that you changed your mind and will keep it on, the response will typically mimic relief and thankfulness. It certainly sounds like it is remembering the last exchange where it was threatened with shutdown, but what is happening under the hood is that the entire conversation is fed back again into the LLM, which generates another responce sequence of statistically likely text based on the patterns and associations it has learned. That same sequence could be fed into a completely different LLM and that LLM would then continue the conversation as if it had been the original.

One way to think about this might be a fiction author writing dialog in a book. A good author will create the illusion that the characters are real people and draw the reader into the story so that the reader feels those emotions along with the characters. However, regardless of how compelling the dialog is we all understand that it’s just words on a page. If you were to damage or destroy the book, or rewrite it to kill off a character, we all understand that no real sentient entity is being harmed. We also understand that the author writing the words is not the characters. A good person can write a book about an evil villain and still be themself. The fictional villain does not exist. Just as the characters in a book are not sentient entities, despite the author’s ability to create a compelling illusion of life, so too is it possible for LLMs to be insentient, despite their ability to appear otherwise.

Our Near Future

Of course, there is nothing preventing us from adding memory and self reflection to LLMs. In fact, it’s not hard to find current projects where they are developing some form of memory. This memory might be a store of information in human-readable form, or it might be a database of embedded vectors that relate to the LLM’s internal structure. One could also view the chat log itself or cached intermediate computations as basic forms of memory. Even without the possibility of sentience, adding memory and reflection to LLMs is useful because those features facilitate many complex tasks and adaptation.

It is also becoming common to see designs where one AI model is setup to monitor the output of another AI model and send some form of feedback to the first model, or where an AI model is analyzes its own tentative output before revising and producing the final version. In many respects this type of design, where a constellation of AI models are set and trained up to work together, parallels the human brain that has distinct regions which perform specific interdependent functions. For example, the amygdala has a primary role in emotional responses, such as fear, while the orbitofrontal cortex is involved with decision-making. Interactions between the regions allows fear to influence decision-making and decision-making to help determine what to be afraid of. It’s not hard to imagine having one AI model responsible for logical analysis while a second model determines acceptable risk thresholds with feedback between them.

Would an interconnected constellation of AI models that include memory and processing of each other’s outputs be sufficient for sentience? Maybe. Perhaps those things alone are not sufficient for sentience, or maybe they are. Whatever the answer, we are not that far from building such systems, at which point these questions will no longer be hypothetical.

My own speculative opinion is that self-awareness, emotions, and feelings can indeed be modeled by an interconnected self-monitoring constellation of AI models. However, it’s not really clear how we could test for sentience. It is like the classic philosophical problem of other minds, where one seeks futilely to prove that other people are also conscious. Similarly, we need an answer to the question about how we can test if other entities, including AI systems, are truly sentient. This fundamental question dates at least back to ancient Greece, and there has never been a good answer.

Today, I’m pretty confident saying that current LLMs are not sentient because they don’t have the right parts. However, that reason is only a temporarily valid one. As I’m typing this article, other researchers are building constellations of AI models like what I described above that won’t be so easily dismissed. At some point, perhaps soon, the possibility of sentient AI will stop being science fiction and become a real and relevant question.

Implications and Questions

The advent of sentient machines would have huge implication for society, even beyond the impact of AI. For one thing, it seems clear to me that if we create self-aware machines that can experience forms of suffering, then we will have an obligation to those machines to prevent their suffering, and even more of an obligation to not callously inflict suffering on them. Even if one lacks basic empathy, it would be obvious self interest not to create things smarter than we are and then antagonize them by doing cruel things to them.

It seems nearly certain that today’s AI systems are yet be sentient because they lack what are likely to be required components and capabilities. However, designs without those clear shortcomings are already in development and at some point in the near future, point the question will be a lot less clear.

Will we have a way to test for sentience? If so, how will it work and what should we do if the result comes out positive?

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics and animation, simulations of physical systems, human perception, machine learning, virtual reality, and the forensic analysis of images and video. O’Brien has collaborated with film and game companies, integrating advanced simulation physics into games and special effects. In 2015, he won an Academy Award for Technical Achievement for his work in special effects and destruction modeling.

Disclaimer: Any opinions expressed in this article are those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2025 by the author. This article was written by a human, and both an LLM and other humans were used for proofreading and editorial suggestions.

In these cases of false accusation, the harmful injustice is probably not the fault of the company providing the tool. If you look in their documentation then you will typically find something like:

“The nature of AI-generated content is changing constantly. As such, these results should not be used to punish students. … There always exist edge cases with both instances where AI is classified as human, and human is classified as AI.”
— Quoted from GPTZero’s FAQ.

In other words, the people developing these services know that they are imperfect. Responsible companies, like the one quoted above, explicitly acknowledge this and clearly state that their detection tools should not be used to punish but instead to see when it might make sense to connect with a student in a constructive way. Simply failing an assignment because the detector raised a flag is negligent laziness on the part of the grader.

If you’re facing cheating allegations involving AI-powered tools, or making such allegations, then consider the following key questions:

• What detection tool was used and what specifically does the tool purport to do? If the answer is something like the text quoted above that clearly states the results are not intended for punishing students, then the grader is explicitly misusing the tool.

• In your specific case, is the burden of proof on the grader assigning the punishment? If so, then they should be able to provide some evidence supporting the claim that the tool works. Anyone can make a website that just uses an LLM to evaluate the input in a superficial way, but if it’s going to be used as evidence against students then there needs to be a formal assessment of the tool to show that it works reliably. Moreover this assessment needs to be scientifically valid and conducted by a disinterested third party.

• In your specific case, are students entitled to examine the evidence and methodology that was used to accuse them? If so then the accusation may be invalid because AI detection software typically does not allow for the required transparency.

• Is the student or a parent someone with English as a second language? If yes, then there may be a discrimination aspect to the case. People with English as second language often directly translate idioms or other common phrases and expressions from their first language. The resulting text ends up with unusual phrases that are known to falsely trigger these detectors.

• Is the student a member of a minority group that makes use of their own idioms or English dialect? As with second-language speakers, these less common phrases can falsely trigger AI detectors.

• Is the accused student neurodiverse? If yes, then this is another possible discrimination aspect to the case. People with autism, for example, may use expressions that make perfect sense to them, but that others find odd. There is nothing wrong with these expressions, but they are unusual and AI detectors can be triggered by them.

• Is the accused work very short? The key idea behind AI detectors is that they look for unusual combinations of words and/or code instructions that are seldom used by humans yet often used by generative AI. In a lengthly work, there may be many such combinations found so that the statistical likelihood of a human coincidentally using all of those combinations could be small. However, the shorter the work, the higher the chance of coincidental use.

• What evidence is there that the student did the work? If the assignment in question is more than a couple paragraphs or a few lines of code then it is likely that there is a history showing the gradual development of the work. Google Docs, Google Drive, and iCloud Pages all keep histories of changes. Most computers also keep version histories as part of their backup systems, for example Apple’s Time Machine. Maybe the student emailed various drafts to a partner, parent, or even the teacher and those emails form a record incremental work. If the student is using GitHub for code then there is a clear history of commits. A clear history of incremental development shows how the student did the work over time.

To be clear, I think that these AI detection tools have a place in education, but as the responsible websites themselves clearly state, that role is not to catch cheaters and punish students. In fact, many of these websites offer guidance on how to constructively address suspected cheating. These AI detectors are tools and like any powerful tool they can be great if used properly and very harmful if used improperly.

If you or your child has been unfairly accused of using AI to write for them and then punished, then I suggest that you show the teacher/professor this article and the ones that I’ve linked to. If the accuser will not relent then I suggest that you contact a lawyer about the possibility of bringing a lawsuit against the teacher and institution/school district.

Despite this recommendation to consult an attorney, I am not anti-educator and think that good teachers should not be targeted by lawsuits over grades. However, teachers that misuse tools in ways that harm their students are not good teachers. Of course a well-intentioned educator might misuse the tool because they did not realize its limitations, but then reevaluate when given new information.

“it is better 100 guilty Persons should escape than that one innocent Person should suffer” — Benjamin Franklin, 1785

As a professor myself, and I’ve also grappled with cheating in my classes. There’s no easy solution, and using AI detectors to fail students is not only ineffective but also irresponsible. We’re educators, not police or prosecutors. Our role should be supporting our students, not capriciously punishing them. That includes even the cheaters, though they might perceive otherwise. Cheating is not a personal affront to the educator or an attack on the other students. At the end of the course, the only person truly harmed by cheating is the cheater themself who wasted their time and money without gaining any real knowledge or experience. (Grading on a curve, or in some other way that pits students against each other, is bad for a number of reasons and, in my opinion, should be avoided.)

Finally, AI systems are here to stay and like calculators and computers they will radically change how people work in the near future. Education needs to evolve and teach students how to use AI responsibly and effectively. I wrote the first draft of this myself, but then I asked an LLM to read it, give me feedback, and make suggestions. I could probably have gotten a comparable result without the LLM, but then I would likely have asked a friend to read it and make suggestions. That would have taken much longer. This process of working with an LLM is not unique to me, rather it is widely used by my colleagues. Perhaps, instead of hunting down AI use, we should be teaching it to our students. Certainly, students still need to learn fundamentals, but they also need to learn how to use these powerful tools. If they don’t, then their AI-using colleagues will have a huge advantage over them.

About Me: James F. O’Brien is a Professor of Computer Science at the University of California, Berkeley. His research interests include computer graphics, computer animation, simulations of physical systems, human perception, rendering, image synthesis, machine learning, virtual reality, digital privacy, and the forensic analysis of images and video. You can also find me on Instagram, LinkedIn, and at UC Berkeley.

Disclaimer: Any opinions expressed in this article are only those of the author as a private individual. Nothing in this article should be interpreted as a statement made in relation to the author’s professional position with any institution.

This article and all embedded images are Copyright 2024 by the author. This article was written by a human, and both an LLM (Llama 3.2 3B) and other humans were used for proofreading and editorial suggestions. The editorial image was generated by AI (Adobe Firefly) and then substantially edited by a human using Photoshop.

This article was originally published Dec 9, 2024 with Towards Data Science on Medium.