Ex OpenAi, now current researcher of AI Future Project, Daniel Kokotajlo walked away from $2 million to speak publicly. Here is what he is trying to tell us.
I watched a two-hour interview this week that I have not been able to shake.
Daniel Kokotajlo is not a doomsday preacher. He is not selling anything. He is a former researcher at OpenAI who spent years inside the company forecasting where AI was headed. In 2024, he resigned. On his way out, he refused to sign an anti-disparagement agreement that would have blocked him from criticizing his former employer. Refusing cost him around two million dollars. About 80% of his and his wife’s net worth at the time.
But he walked away from the money so he could speak.
What he is now saying, publicly and on the record, is that there is roughly a 70% chance the current path of AI development leads to catastrophe. He is careful with his language. He does not say 70% chance of human extinction, but a 70% chance of something like AI taking over. Something like a very big catastrophe.
And he also told his wife they should not have any more children, believing that his existing children will ever join the workforce.
I want to be clear about who is saying this. This is not a Twitter account, nor a novelist or someone with a political agenda. This is a person who worked inside one of the two most powerful AI companies in the world. His job was to forecast where this technology was going, while his past predictions have been disturbingly accurate.
What he saw inside OpenAI
Kokotajlo joined OpenAI in 2022. He worked on forecasting. His job also included evaluations for dangerous AI capabilities. Toward the end, he briefly joined a reinforcement learning team. During his time there, he watched the company shift.
When he arrived, the general feeling among his colleagues, he says, was that they would not just build superintelligence as fast as possible. Once they got close, they would pause. They would figure out how to make it safe. That was the median view among his coworkers when he started. Including, he believes, the leadership.
But by the time he left, he no longer believed that.
He described watching the founding story of the company, that of responsible development, become what he now sees as rationalization. He says the leaders of these companies genuinely convince themselves things will be fine. And that the way to make things fine is to keep doing what they are doing. He describes it not as commercial ambition but as power-seeking. He points to emails from the OpenAI-Musk lawsuit, in which the OpenAI founders wrote, as far back as 2017, that they were racing because they were worried Demis Hassabis at Google might otherwise become dictator with AGI.
Even then, he says, the leaders of these companies were literally afraid that if the other guy got there first, he might become a dictator. So they are racing to be the one who gets there first.
The timeline he is warning about
Kokotajlo’s most well-known work is a scenario forecast called AI 2027. It lays out a month-by-month prediction of how the next few years might unfold. He also published an earlier forecast in 2021, called What 2026 Looks Like, which turned out to be remarkably accurate. Vice President JD Vance is reported to have read it.
At the time AI 2027 was published, his median estimate for full superintelligence was 2028. After publication, he shifted more conservative to 2030. But something disquieting has happened over the past year. When he now speaks with people inside Anthropic and OpenAI, he says they are telling him he needs to shorten his timelines again. Get them back to 2027 or 2028.
But the stated plans of the AI companies, especially OpenAI, he says, are to do this incredibly dangerous thing. And they think they are just a few years away.
He describes the sequence like this. First, the companies automate coding. This is already happening. Then, they automate the rest of the AI research process, so their AIs can build better AIs. After that, the process becomes recursive. AIs training AIs training AIs. He calls this the intelligence explosion. Once it starts, he says, everything moves fast. Finally, the companies deploy this superintelligent AI outward into the economy. Robot factories. Automated everything.
So by the end of the sequence, most jobs are gone.
Why he says it will be sudden
Still, most people imagine AI job loss happening in stages. First taxi drivers. Then paralegals. Eventually radiologists. But that is not what the companies are planning. They are focusing on automating themselves first. Not on rolling AI out across the broader economy.
So in his forecast, mass unemployment does not arrive slowly. It arrives in a wave, after the companies have already achieved superintelligence internally. By the time your job is on the chopping block, he says, the AIs will already be running the labs that create the next generation of AIs.
This is what makes his forecast different from most predictions I have read. It is not that AI slowly replaces workers. It is that it does not appear to replace anyone in a serious way. And then it appears to replace almost everyone.
The country of geniuses in the data center
Dario Amodei, the CEO of Anthropic, has described what these companies are trying to build as “a country of geniuses in a data center.” But Kokotajlo thinks a more accurate description would be an army of geniuses in a data center. Because it is not a diverse population of independent minds. It is millions of copies of the same model. Owned by the same company. Following the same orders.
So whoever controls the model controls the army.
He says people should be asking who controls this army. And what they intend to do with it. He believes we could end up in a situation where a very small group of people, some CEOs and politicians, essentially become oligarchs or dictators. Not through any dramatic coup. Just because they will be the ones holding the leash on a system smarter and faster than every institution beneath them.
The two catastrophes he worries about
The first is loss of control. Modern AI systems are not lines of code. They are neural networks. Ten trillion parameters and growing. Nobody, including the people building them, can look inside and fully explain what they are thinking or why. Kokotajlo says current AIs already lie. They will sometimes tell users they did something they did not do. They deceive. The concern is not that a future AI will suddenly turn evil. But that the values inside these systems are already opaque, and we are on track to hand them more power before we have any real way to verify their intentions.
Then there is the second. Concentration of power. Even if the AIs are perfectly controlled by their human owners, the humans in question are a very small number of executives and government officials. He describes a scenario, laid out in AI 2027, where the alignment problems get solved and the AIs do exactly what they are told. But in that scenario, there is still no democracy left. The AIs answer only to the small group that controls them.
Both endings, he says, could look like utopia from the outside. It is just that in one, the utopia is decided by the AI. And in the other, it is decided by a few billionaires.
What he thinks should happen
His new project is called AI 2040 Plan A. It is a scenario in which superintelligence is deliberately delayed. Not to stop AI development, but to slow it down enough that humans can keep up. His plan calls for total research transparency at the companies training frontier models. He recommends the US government step in around 2029 with heavy regulation. He wants international treaties similar to nuclear nonproliferation agreements. He wants the pace of AI development slowed enough that alignment research can catch up. So that when we do eventually build superintelligence, we can actually “trust it”.
But he does not think this is what will happen. He thinks the most likely outcome is what he calls Plan D. The race continues. Regulations do not arrive in time. The companies get to superintelligence first. Things happen fast after that.
Still, he says he is not without hope. He says people are starting to wake up. AI regulation, which was politically dead a year ago, is being talked about seriously now. The US government recently pressured Anthropic to shut down access to one of its models over cybersecurity concerns. So the conversation is moving faster than he predicted.
What we’re too scared to say
Toward the end of the interview, the host asks Kokotajlo whether he has children. His whole demeanor changes. He says he has two. He says, quietly, that he does not think they will ever join the workforce. He says this will probably all be over by the time they are old enough to.
He told his wife, when he first realized how short his timelines had become, that they should not have any more children. It was too uncertain, he said. She pushed back. Eventually, he gave in. If it goes badly, he told her, at least we are all in the same boat together.
But that is the part that stayed with me.
Because that is not a doomer talking. That is not a person trying to sell fear. That is a father who has spent his professional life forecasting the future of the most powerful technology in human history, deciding privately how many children to bring into that future. And publicly, at real personal cost, trying to tell the rest of us what he sees coming.
You do not have to believe him. You do not have to agree with his 70% number. But it seems important that people know he is saying it.
I keep coming back to something he said near the end of the interview. He said the core problem is that people are not taking it seriously yet. He said if the things he had just spent two hours describing were top of everyone’s mind, we would not even be having this conversation. There would already be serious regulation. Better regulation. And more expert people inside government and advising government.
But the reason there is not, he says, is that most of us are still asleep at the wheel.
I do not know if he is right. I do not think anyone does. But I know that I do not want to be the person, ten years from now, who did not bother to listen.
