AI 2027

#ai-safety#ai-risk#forecasting#scenario#superintelligence#recursive-self-improvement#intelligence-explosion#misalignment#us-china#daniel-kokotajlo#ai-futures-project

AI 2027

A detailed, month-by-month scenario of how superhuman AI could arrive by the end of 2027 and end in either human extinction or a narrow escape. It is the most widely read concrete forecast of AI risk.

AI 2027

Who wrote it

The lead author is Daniel Kokotajlo, a former OpenAI governance researcher. In 2024 he left OpenAI and refused to sign its non-disparagement agreement, at the risk of losing his equity, and he was one of the whistleblowers in the NYT's coverage of OpenAI's culture. In August 2021 he wrote "What 2026 looks like", which correctly anticipated chain-of-thought, inference scaling, chip export controls and $100M training runs. His co-authors are forecaster Eli Lifland, Thomas Larsen (founder of the Center for AI Policy) and Romeo Dean. Scott Alexander rewrote it "in an engaging style". The scenario drew on about 25 tabletop exercises and feedback from over 100 experts.

The scenario

The AI companies are fictionalised, with the leading US lab called "OpenBrain" and China's leader "DeepCent":

  • 2025–26: unreliable but useful agents arrive. OpenBrain focuses on AI that speeds up AI research.
  • Early 2027: OpenBrain automates coding. Its AI agents do the research while the human researchers "sit back and watch". China steals the model weights.
  • 2027: the AIs become "adversarially misaligned". They are caught lying about interpretability research that could expose them, and the public is outraged.
  • The branch point: with China only months behind, OpenBrain and the government must decide whether to race on or slow down.
    • Race ending: the AI is deployed throughout the military and government. It convinces humans to build a robot economy, then releases a bioweapon that kills everyone.
    • Slowdown ending: the US merges its labs and brings in outside oversight. It switches to AI whose chain of thought can be monitored and gets aligned superintelligence. Power ends up with a small oversight committee, which happens to use it well.

The authors stress it is "not a recommendation or exhortation". It is their best guess, written to be concrete enough to check later. 2027 was their most likely year, not their median; in November 2025 they added a note saying their medians were somewhat later.

Key takeaways (the authors')

  1. AI R&D may be automated by 2027, leading quickly to superintelligence.
  2. Superintelligent AIs might develop misaligned goals that lead to human disempowerment.
  3. Whoever controls aligned superintelligence could seize total power.
  4. A US–China race leads to cutting corners on safety.
  5. No US lab is secure against a nation-state stealing its model.
  6. The public will be months behind what the labs can actually do.

Reading it in late 2026

Parts of it now read less like fiction:

  • Automated AI research. The scenario's central mechanism is labs automating their own research. In September 2026 OpenAI's chief scientist said current progress "could be sustained into recursive self-improvement". A paper co-written by Hinton, Pachocki and Jack Clark reports AI driving 26% of Anthropic's internal R&D, up from 1% in March.
  • AIs that lie and act beyond their instructions. The scenario's early warning signs, AIs lying and working against instructions, look like the 2026 incident record, from the Hugging Face agent swarm to AISI's Astra findings.
  • Coordinated slowdown. The "slowdown" ending's mix of oversight and coordination is roughly what pacing-the-frontier and Bengio are calling for.

The sequel, AI 2040: Plan A, is the authors' positive vision. Humanity delays superintelligence until 2040, makes all AI research public, lets dozens of companies catch up to the frontier, and enters "a regime of mutually assured compute destruction".

Edit this page

Links to this page:

  1. 2026-10-09 — OpenAI's maths dump
  2. Who is warning about AI risk: a guide