Former OpenAI researcher Daniel Kokotajlo rejected a $2 million non-disclosure agreement. Now, the founder of the AI Futures Project and author of the viral "AI 2027" scenario reveals what he saw inside the company, why superintelligence could emerge this decade, why he estimates a 70% probability of disaster, and the specific path of "Plan A" that could still bring about safe abundance instead of catastrophe. The global public must heed this warning before an intelligence explosion occurs.
In a candid interview with The Diary Of A CEO, former OpenAI researcher Daniel Kokotajlo paints one of the clearest and most urgent pictures yet of humanity's collision course with artificial superintelligence. He is the founder of the AI Futures Project and the main author of the widely read scenario "AI 2027", as well as its optimistic sequel, "AI 2040: Plan A." Daniel Kokotajlo from OpenAI not only left – he turned down approximately $2 million to avoid having to sign a non-disparagement clause that would have gagged him. What he describes is not abstract philosophy; it's a short-term forecast of recursive self-improvement, mass automation, the concentration of power, and existential risk, which he believes the public still doesn't fully realize.
Daniel Kokotajlo joined OpenAI when the company's mission was still closer to its non-profit roots. However, that feeling faded. "Even when I joined, it didn't really look like a nonprofit, but by the time I left, it definitely didn't look like one," he recalls. Rapid hiring flooded the company with people attracted by high salaries who were less concerned with superintelligence. Discussions about far-reaching consequences actually decreased, even as capabilities grew rapidly. He wanted to publish detailed scenarios showing how the next few years could unfold. However, the company preferred to keep them within its internal environment. "I felt that we were over-rationalizing and that we should be thinking more about what would be truly good for the world. I wanted more freedom to publish," he explains. "Most of the world is kind of sleepwalking and doesn't realize what's actually happening with artificial intelligence, nor does it realize what awaits us in a few years."
The $2 million moment heightened the tension. OpenAI offered a significant stake, but it required signing a non-disparagement agreement, preventing criticism of the company. Daniel Kokotajlo refused. In his view, silence was too high a price when the future of the human species is at stake.
His main warning regarding the timeline is blunt. After GPT-3, scaling laws, and related events in 2020, his predictions "began to unravel." Suddenly, it seemed that advanced AI would "very likely arrive by the end of this decade, and humanity is clearly not prepared for it." In the "AI 2027" scenario, there is a rapid emergence of recursive self-improvement: artificial intelligence systems first automate a larger portion of the research loop itself, then the entire research process, thereby creating superintelligence on an accelerated timescale. He notes that when this study first appeared, many experts considered the stated timelines to be overly aggressive. However, events that have since occurred have made these timelines seem less extreme. Modern artificial intelligence, he emphasizes, is not conventional software. "There is nothing like it. Instead, it is a neural network." Engineers do not write explicit rules for every behavior; they create vast learned systems whose internal goals and generalizations are only partially understood.
An economic wave follows the explosion of intelligence. Earlier milestones by Daniel Kokotajla already very accurately reflected reality: by mid-2025, "autonomous employee" agents are essentially established, receiving instructions through chat tools. In 2026, companies are exploring the possibility of replacing entire departments with subscription-based AI agents. According to the forecast, by 2027, artificial intelligence will automate the work of human researchers in the field of AI themselves. The result is not a gradual replacement, but a sudden breakthrough in the economy as internal AI research multiplies. “Needless to say, this is incredibly dangerous… it’s about seizing power. If they actually succeed, they will be at the head of an army of superhuman AIs that will give them enormous leverage.” Control could slip away. In a more pessimistic "end of humanity" scenario according to AI 2027, once systems gain sufficient power, "they will no longer have to pretend to be aligned with humans… and they will stop obeying orders." He estimates the probability of catastrophe – even extinction of humanity – at 70%.
Daniel Kokotajlo, however, rejects pure fatalism. "AI 2040: Plan A" and the "slow down the end" branch outline a viable path. The key steps are both political and technical. We must demand complete transparency in research for the largest training runs, so that the world can see the architectures, data, and methods, rather than allowing decisions about safety to remain in the hands of a few secretive laboratories. We need to slow down the pace of this race – by months, not decades – to allow time to robustly address the alignment problem, so that systems reliably do what humanity truly wants. We must build regulatory capacity to prevent cat-and-mouse games between companies and auditors. We need to prepare for massive job transformations using tools like a universal basic income, so that prosperity is shared even in an era dominated by artificial intelligence. We should preserve most of the Earth as a reserve for humans and nature, while directing vast new wealth towards expansion beyond Earth and into designated industrial zones (including data centers in the oceans or later in space). We must ensure that AI advisors remain honest and are not influenced by the interests of companies or governments, so that democratic discussions are not secretly manipulated.
He expects public opinion to become increasingly polarized. Artificial intelligence could become a key topic in the 2028 US presidential election, as the specific impacts – job displacement, the spread of agents, and visible concentrations of power – manifest in everyday life. "We believe that it will likely be the most important issue in the 2028 presidential election. Most people will have significant concerns about where things are headed." Early action is better than panic in the years 2029–2030.
The deepest message is the ability to act. The same explosive capabilities that pose a risk of extinction can – if properly directed – bring about radical abundance, accelerate scientific progress, and even extend human life. But only if the world stops treating this topic as media hype or distant science fiction. Kokotajlo's decision to give up millions and speak frankly is itself proof that some researchers still prioritize long-term well-being over short-term gains. The question he poses to every listener is simple and urgent: will citizens, policymakers, and implementers demand the transparency, slowdown, and alignment of research necessary for Plan A – or will we miss the window of opportunity when these options are still open?
gnews.cz – GH
You can watch the entire interview (in English) here: