Transcription
LLM psychosis is going to be a really hot topic in 2026. We already see it coming up in lawsuits that model makers are facing as loved ones alleged that people who committed violent acts were somehow induced to do that by artificial intelligence. It's going to get into the workplace next year. And the reason why is that people who you would think are very sober, very levelheaded still show evidence very publicly of LLM induced psychosis.
The most prominent example recently has been David Buden, a former director of engineering at Google Deep Mind, now the founder and CEO of the company Pingu, who bet publicly $10,000 he could solve Navier Stokes. Navier Stokes is a fluid dynamics equation. And the long and the short of it, if you're not mathematical, is that we cannot perfectly prove how fluids move using equations. We tend to approximate them at very high fidelities because their movements are so complex. Solving Navier Stokes and showing how they work mathematically has been a millennium prize effort. So it carries a prize of a million dollars. David absolutely went out and published a bunch of chat GPT 5.2 to prodriven equations and what he called a lean proof over the weekend of December 20th and then claimed that by December first he would publish a full proof of Navier Stokes. Now mathematicians looked at this, people much smarter than me looked at this and everyone is convinced looking at his initial work and his lean proof that David has been suffering from LLM induced psychosis and is perhaps the most prominent recent example of that issue.
But I got to tell you, he's not the only one. I have seen symptoms of this in people that I know over the course of 2025. And it's going to become more and more concerning in the workplace because you're going to need to know that the human that is making the decisions while they may engage with AI has not had their brain hijacked by AI. In this case, David's brain seems to be convinced by chat GPT that he's close to solving Navier Stokes when some of our most prominent mathematicians namely Terrence Tao are not even convinced it's solvable. It may not be sub subject to a single smooth equation.
So, with that in mind, here are my tips for you to avoid LLM psychosis.
Number one, please, please, please ask your LLM to be adversarial with you regularly. That was one of the things that people noticed in the prompts that David Budden shared. Even though he's asking the AI to check his work, he's not doing so in an adversarial way. In fact, he's doing so in a confirmatory way. That is a classic symptom of LLM psychosis. When you want the AI to agree with you, you tell you tell it to check your work, but you don't really want it to check your work. You want it to tell you what you want to hear. In this case, he wants to hear that Navier Stokes has been solved. And so, he wants the LLM to show that that's the case.
Number two, do not assume that just because you have an LLM in your pocket or on your laptop, you are suddenly a budding cutting edge scientist or mathematician who can do things that the brightest minds on the planet have not been able to do ever. I know we talk about how smart these systems are, but you still need to be a very smart person with deep domain experience to validate and check scientific hypotheses, mathematical theorems, etc. And that actually goes for the rest of work too. If you if you get told by chat GPT that there's a better way to invent and install solar panels, if you don't have the domain expertise in solar, you cannot know it's correct. and Ched GPT telling you it's correct isn't worth a whole lot. And so one of the things that we need to start seeing more of is an awareness that even though we can expand our spans dramatically with AI, our domain expertise matters more and more and more because we are going to be the ones that need to check these things for sanity. We are going to need to be the ones that say this actually works in the real world or it does not. And increasingly, not just with David Button, with many others I've met, I can count probably a dozen that I know of, I see instances where people are perhaps not in full LLM induced psychosis. There's no danger to loved ones, anything like that, but they are not able to distinguish between their own expertise and chat GPT's expertise, and they have an inflated sense of what they are capable of that is not correct. Yes, you can do a lot more work with an AI, but it comes from your own expertise and your own ability to actually get work done and know what good looks like. And this is why I keep emphasizing that engineers are not out of jobs. You can get LLMs to write lots and lots of terrible code. That's cheap and easy. It is very hard to get LLMs to write code in modules that pass evals within a structure that works at a scaled production system. That takes engineering. And that is why I firmly believe that we will not have Betty and HR vibe coding a CRM or vibe coding an HR information system in 2026. We are going to have traditional software providers that are extending and personalizing software like HR information systems for people like Betty, but that will be done by professionals who have deep expertise. And so as much as I love vibe coding and I think there's it's a tremendous unlock for engineers, it's a tremendous productivity lock internally for companies is different from saying anybody can make anything without having domain expertise. That's just not true. You need the domain expertise to actually be able to successfully accomplish meaningful work. So that that's that's the other one that I want to call out.
The the third thing that I would call out, so like we talked about the fact that you need to ask for disconfirming information. and ask your AI to get adversarial with you. We talked about the fact that you need to not overstate your own domain expertise. The third thing that I want to call out is you need to submit to a jury of your peers. If your peers as a whole in your domain think you are out to lunch and think you are incorrect, a symptom of psychosis is to say, "No, me and AI are right. Y'all are wrong. Y'all are the ones that don't have this figured out. The AI and I have it figured out." That's LLM induced psychosis right there. You may not be a danger to yourself and others, but you're not entirely well in the head. Because if your peers who have deep domain expertise strongly disagree with you and like almost every one of them does, then it is a sign for you that you are missing something. And if you cannot hear other humans, you are going to be in trouble in 2026.
One of the signs of stable leadership in 2026 is going to be the ability to know when to turn the laptop off, when to shut chat GPT down, turn all the recording devices off, and have a conversation, talk to a human, make a business decision, understand what really needs to be done. Stable leaders are going to be able to do that, and people who are unstable are going to need AI with them all the time in order to make any kind of decision like that. and they will be very disagreeable to work with because they will say, "I'm right and AI is right and you're wrong all the time." And that's actually one of the ways that we know that David Buden is probably not solving Navier Stokes in 3 or 4 days because all the mathematicians that looked at the lean proof were like, "H, this looks a little bit shaky." I'm not a mathematician. I'm not saying that I looked at it cuz I don't believe I have that expertise. I'm looking to a jury of my peers. I'm looking to people who know about science and math more than I do. And when they all are like, "This looks sketchy." I'm like, "It's probably sketchy." You need that degree of common sense. You cannot substitute for common sense like that. You need the ability as a leader to know when AI is not going to be helpful. And that's true not just as a leader, but for all of us, whether at work or in our personal lives. And so, as much as I think it's likely that we will eventually have LLM induced psychosis in the DSM5 as a recognized psychiatric disorder, we should not wait for that. And businesses are going to start to test leaders probably quarterly to ensure that leaders are not under undue influence by AI because if you are, you're risking your whole business and it's just not safe. We are not at a point where it is a safe or good thing for a human to be unduly influenced by AI as they make business decisions. And increasingly I see that LLM induced psychosis is not limited to people who are on the edges of society. People who like David Bud, our CEOs, our founders, prominent leaders, can still fall victim to this idea that them plus the AI equals some sort of incredibly intelligent being that beats everybody else. That's just not the way it is. You plus AI is just you with a tool and you need your colleagues to work with you to get meaningful work done. And as cool as AI is, and as much as it's transformational, that is going to remain true in 2026. and businesses are just now at the beginning of figuring out what it looks like to actually get good testing in on LLM psychosis. We we would have to write those tests. I'm going to start to do some more thinking in that direction because I think it's going to be one of the key leadership traits that we will test for and verify and think about as we move forward. It won't just be can you use AI. It will be can you use AI and not go crazy. So don't go crazy. Your AI is just a tool. If your peers all think you're out to lunch, you're probably out to lunch. And uh don't try to solve Navier Stokes.