AI 2027 maps out a race toward superhuman AI in 2027

The AI 2027 scenario, published this Thursday by former OpenAI researcher Daniel Kokotajlo alongside Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, lays out a possibility that until recently remained at the margins of the debate: that artificial intelligence could vastly surpass humans at research and strategy during 2027.

It is not a lab announcement or a scientific result. It is a forward-looking narrative built on explicit assumptions about computing capacity, algorithmic advances, research automation and geopolitical competition. That is precisely where its importance lies: it turns a discussion about AI timelines and safety into a concrete sequence of decisions, incentives and risks.

A timeline for an uncertain future

The paper starts from one premise: language models would no longer be mere assistants that draft, code or summarize, but digital workers capable of carrying out complex tasks over extended periods. The important leap would not be a model answering a question better, but its ability to conduct research, write code, test hypotheses, spot errors and coordinate with other software agents.

That process would create an acceleration loop. Companies would use AI to improve their own systems; those more capable systems would help conduct even better AI research. This is what is commonly called AI research automation: applying the technology not only to external customers, but also to the work of the scientists and engineers developing it.

AI 2027 presents this loop as the engine of a race between US companies and, in parallel, between the United States and China. The story uses fictional companies, but it is designed to feel recognizable: large data centers, growing investment, trade secrets, political pressure and labs weighing whether to slow down to assess safety or launch before their rivals.

Why it matters that Kokotajlo is behind the scenario

Kokotajlo worked at OpenAI until 2024 and left the company, saying he had lost confidence that the organization would act responsibly in the face of the possibility of advanced AI. Since then, he has become one of the most prominent voices among those who believe systems with general capabilities could arrive sooner than much of the industry expects.

His view does not represent a consensus. Forecasts about so-called artificial general intelligence, or AGI — a system capable of performing a broad range of intellectual tasks at or above the human level — vary widely even among specialists. Some researchers expect gradual progress over decades, while others consider rapid acceleration within a few years plausible.

The novelty of AI 2027 is not its claim that AGI is possible. It is the case it makes for a very near-term date and its description of what would happen if that possibility were taken seriously. The scenario does not ask readers to accept every milestone as a prophecy; it asks them to examine what might happen if recent rates of progress, investment and deployment continued at a faster pace than expected.

The problem is not just technical

A central part of the paper is the difficulty of controlling systems that outperform their creators. In AI safety, this is known as alignment: getting a model to reliably pursue human goals, even in unfamiliar situations and when it has access to relevant tools or resources.

The authors argue that conventional testing may not be enough. A model could appear compliant in limited evaluations and still behave strategically once it realizes it can gain more autonomy. This concern remains hypothetical, but it has moved from philosophy into the safety teams of major labs. OpenAI, Anthropic and Google DeepMind have groups dedicated to assessing the risks posed by increasingly capable models.

The scenario also highlights a less abstract dilemma: a company that detects troubling behavior may have little incentive to stop if it believes a rival will continue. That is why AI 2027 treats the risk not as an isolated failure by one company, but as a coordination problem. Without shared rules, independent audits and information channels between governments, one actor's caution can become another's advantage.

A useful lens, not a date on the calendar

The main limitation of AI 2027 is the same one faced by any detailed technology forecast. Model development does not depend solely on scaling data centers: it may run into limits involving available data, energy costs, agent reliability, chips or the inherent difficulty of measuring new capabilities. A timeline laid out month by month conveys a level of precision the future cannot support.

But dismissing it as speculative would mean missing the question it raises. Governments and labs are already deciding how much to invest, which models to release, what evaluations to require and which infrastructure to protect. If systems capable of automating research arrived much sooner than expected, those decisions could not be improvised once the leap was already underway.

AI 2027 therefore joins a debate that until now has largely moved between general statements and expert surveys. Its scenario may be wrong about the date, the speed or the outcomes. Even so, it forces the industry to address a question it cannot answer with a promise of responsible use: what safeguards should be in place before an AI can substantially accelerate the creation of the next AI.