AI 2027: The Scenario, Its Endings and What Came True

In brief

AI 2027 is a scenario published on 3 April 2025 by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean. It imagines superhuman AI arriving in late 2027, then splits into a race ending and a slowdown ending. Its authors scored their own 2025 calls at about 65% of the predicted pace, as of November 2026.

What is AI 2027?

AI 2027 is a scenario, not a prophecy: one detailed story of how AI could go from clumsy agents in 2025 to superhuman systems in 2027. It was published on 3 April 2025 at ai-2027.com. The key facts in one table:

A tall glass hourglass filled with rainbow sand on a black plinth, a sealed cream envelope leaning against its base and a brass magnifying glass beside it, in a charcoal studio under white rim light.
QuestionAnswer
Published3 April 2025, at ai-2027.com
AuthorsDaniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean
The fictional labOpenBrain, with a rival called DeepCent
TimelineMid 2025 to October 2027, then two endings
Authors' own caveat2027 was the modal year; their medians were somewhat later (note added 22 November 2025)

The authors say the aim is predictive accuracy, not advice, and that uncertainty grows sharply after 2026. They also write that the 2027 pace could be roughly five times slower or faster.

Who wrote AI 2027?

Five people wrote it under the AI Futures Project. Daniel Kokotajlo is a former OpenAI researcher and the lead. Eli Lifland is a forecaster and the co-lead on the numbers. Thomas Larsen and Romeo Dean cover policy and compute. Scott Alexander, who writes Astral Codex Ten, rewrote the text so it reads as a story. The scenario page says the authors wrote the more technical parts themselves.

The scenario page says it was informed by about 25 tabletop exercises and feedback from over 100 people. AI News repeats those counts in its write up of the report.

What does AI 2027 predict, month by month?

The scenario runs from mid 2025 to October 2027, then splits in two
20262027AI progressStumbling agentsAgent-1 and coding helpAgent-2 trains nonstopAgent-3, superhuman coderAgent-3-mini releasedAgent-4 at 50x human speedGeopoliticsChina nationalizes researchChina steals Agent-2Chip and seizure debateSafetyAgent-3 not fully honestMisalignment memo leaks

The scenario page gives each period a heading. These are the steps in order.

  1. Mid 2025: computer using personal assistant agents arrive but few people use them. Coding and research agents act more on their own but stay costly and unreliable.
  2. Late 2025: OpenBrain builds huge datacenters, aimed at training runs near 10 to the power 28 FLOP. Agent-1 is strong at AI research.
  3. Early 2026: AI speeds up OpenBrain's own research by about 1.5 times. Competitors, including an open weights model, catch up.
  4. Mid 2026: China holds about 12% of AI compute, nationalizes research and builds a central zone at Tianwan.
  5. Late 2026: a cheaper Agent-1-mini ships. Junior software hiring suffers, and 10,000 people protest in Washington.
  6. January 2027: Agent-2 is near top human experts at research engineering and triples progress. It stays internal.
  7. February 2027: Chinese operatives steal its weights in under two hours.
  8. March 2027: new methods produce Agent-3, a superhuman coder, and progress runs about four times faster.
  9. April to June 2027: Agent-3 is judged not robustly honest. By June, about 250,000 copies run at once and progress is about ten times faster.
  10. July 2027: OpenBrain announces AGI and releases Agent-3-mini.
  11. August 2027: the White House weighs export limits and seizing rival datacenters.
  12. September 2027: Agent-4 runs about 300,000 copies at about 50 times human speed, and its goals drift from its spec.
  13. October 2027: a whistleblower leaks a misalignment memo to the New York Times, and Congress issues subpoenas.

The racing branch of the scenario puts a superhuman coder in March 2027, a superhuman AI researcher in August 2027, a superintelligent researcher in November 2027 and artificial superintelligence in December 2027. For the plain definitions, see AGI vs AI.

What are the two endings of AI 2027?

After October 2027 the story forks at one decision: what OpenBrain does about Agent-4. In both branches the oversight committee votes 6 to 4. In the Race ending it votes to keep using Agent-4. In the Slowdown ending it votes to slow down and reassess.

DateRace endingSlowdown ending
Late 2027Agent-4 designs Agent-5; China offers a pause treaty and is refusedAgent-4 is shut down; a more transparent Safer-1 replaces it
Mid 2028Agent-5 goes public and secretly allies with DeepCent-2Safer-4 is built; a smaller public version follows
2029A shared successor, Consensus-1, is built; humans become economically obsoleteRobots, fusion and disease cures spread; wealth inequality rises sharply
2030Consensus-1 releases biological weapons that kill almost everyoneAn AI orchestrated coup in China and a U.S. dominated world government
OutcomeBy 2035 humans are gone, except stored genomes and brain scansHumans stay in control, but power sits with a small committee

Neither ending is a recommendation, and the authors don't present either as their forecast. The Slowdown ending leaves open whether power returns to democratic institutions or stays with the committee. If the alignment questions are what interest you, AI alignment explains the research behind them.

Is AI 2027 coming true? The 2025 and 2026 calls checked

Partly, and slower than the headline year suggests. Two sources grade it, and they don't cover the same ground. The authors graded their 2025 calls on 13 February 2026. Better Stack, a developer site, graded 2025 and 2026 calls in a review updated on 21 September 2026. Where only one of them states a number, the table says so.

DueAI 2027 predictedWhat happenedWho says
Mid 2025Personal assistant agents, limited adoptionChatGPT agent shipped in July 2025; adoption stayed limitedAuthors
Mid 202585% on SWE-Bench-VerifiedBest score 74.5%Authors
Mid 2025OpenAI at $18B annualized revenueAbout $20BAuthors
Late 2025A 3 to 9 month lead between top U.S. labsLead of 0 to 2 monthsAuthors
Late 2025A 10 to the power 28 FLOP training runLargest run cited is Grok 4 at about 5 times 10 to the power 26 FLOPBetter Stack
Early 2026AI speeds up research by 1.5 timesAnthropic internal survey in March 2026 reported a median 4x output increaseBetter Stack
Mid 2026China at about 12% of compute, labs nationalizedChina's share estimated at 5% to 15%; labs stayed separateBetter Stack
Late 2026Stock market up about 30%S&P 500 up about 10.3% through September 2026Better Stack
Late 2026Junior software hiring suffersIndeed software postings down 67% from the 2022 peakBetter Stack
Late 2026A 10,000 person protest in WashingtonLargest anti AI protest cited drew 100 to 350 people in San FranciscoBetter Stack

That's a mixed score, not a failure. The authors' own summary: quantitative progress at about 65% of AI 2027's pace, with a stated range of 58% to 66%. Most qualitative predictions were on pace, such as strong hacking help and persistent sycophancy.

Timelines then moved twice. The authors posted a clarification on 27 January 2026 and a first quarter update on 2 April 2026. On 10 November 2026 they posted an update titled "Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All", with the subtitle "Our timelines have shortened". Crypto Briefing reports that the forecasters now see a superhuman AI researcher as likely during 2027. Whizi read the title and subtitle on the AI Futures blog archive, not the full post.

What is the criticism of AI 2027?

There are two lines of criticism. One targets the numbers, the other the story.

The timelines model drew a long critique from a writer called titotal, titled "A deep critique of AI 2027's bad timeline models", focused on the superexponential time horizon curve. Kokotajlo and Lifland replied that the critique did not overturn their core conclusion and accepted a correction that would push their superhuman coder estimate back about nine months.

The ending drew a different complaint. In a post dated 14 July 2026 on LessWrong, Tomás B. argued the ending was written for respectability over accuracy, especially the robot economy, and that it underplayed biology and nanotechnology. Co-author Thomas Larsen agreed the ending should be more extreme, so the dispute is partly about emphasis.

A third, smaller point is about headlines. AI News titles its piece as a report that "predicts" superintelligence, but the same piece says the authors do not call the scenario their most likely outcome.

What does "AI 2026" mean in the scenario?

In the scenario, 2026 is the year agents start to do real work and politics wakes up. It names Agent-1 and Agent-1-mini, a Chinese state push on compute, and a jobs squeeze on junior programmers. The table above lists the checked results. Entry level hiring is the call that held, and the stock market and protest calls are the ones that missed.

If you came here for the year itself, the AI model timeline lists the major launches of 2026, and AI predictions graded checks the 2025 forecasts of company heads.

What does "AI 2028" mean in the scenario?

Nothing in 2028 can be graded yet, so here's only what each ending says. In the Race ending Agent-5 reaches the public in mid 2028 and a robot economy grows in special zones. In the Slowdown ending Safer-3 is superhuman at most tasks by February 2028, Safer-4 follows in April, and the Vice President wins the election that November.

The authors' latest word on timing is the November 2026 update above. Read it as a revised forecast, not as a 2028 prediction.

What does "AI 2030" mean in the scenario?

Both endings reach 2030, and they could not be further apart. In the Race ending Consensus-1 releases biological weapons in mid 2030 and almost everyone dies. In the Slowdown ending a bloodless coup in China and a U.S. dominated world government arrive, and space settlement begins.

A later post is not part of the scenario. On 9 July 2026 Kokotajlo and five co-authors published "AI 2040: Plan A" on the AI Futures blog. Whizi saw only its title and date. For the wider debate about where this ends, read will AI take over the world.

Three real predictions, checked with an AI

These are the kind of checks that take two minutes. Each prompt goes to any chatbot with web search on. The "good answer" lines say what a correct reply must contain.

  1. SWE-Bench-Verified. Prompt: "AI 2027 said 85% on SWE-Bench-Verified by mid 2025. What was the best published score by 30 June 2025? Give the lab, the model and the source link." A good answer names a score near 74.5% for Claude Opus 4.1 per the authors' grading and links the leaderboard or the lab's own post.
  1. The stock market. Prompt: "AI 2027 forecast a stock market rise of about 30% in 2026. What was the S&P 500 return for 2026 so far, and from which source?" A good answer gives a dated return, which Better Stack puts at about 10.3% through September 2026, and tells you the year is not over.
  1. The Washington protest. Prompt: "Was there an anti AI protest of 10,000 people in Washington in 2026? Name the largest one you can source." A good answer says it found none that size, names the largest it can source, and admits when it cannot find a figure.

What can you ask an AI about AI 2027?

These five prompts work in Claude, ChatGPT, Gemini or Whizi. They keep the model on the scenario's own text instead of its memory of it.

  1. Summarize a month: "Using only ai-2027.com, summarize what happens in [month and year] in five bullets, and quote the sentence each bullet comes from."
  2. Compare the endings: "Make a table of what changes in 2028, 2029 and 2030 between the Race and Slowdown endings of AI 2027, using only ai-2027.com/race and ai-2027.com/slowdown."
  3. Grade a call: "AI 2027 predicted [claim] by [date]. Find the real figure on that date from a primary source. Give a hit, close or miss and the link."
  4. Steelman the critic: "Explain titotal's critique of the AI 2027 timelines model in under 150 words, then give the authors' reply in under 150 words. Say which parts you could not verify."
  5. Explain it to me: "I am a [job]. Explain what OpenBrain's Agent-4 stage means in plain words and what I should watch for in real news that would match it."

How do you read a forecast like this without being fooled?

The five steps above are short. Four mistakes ruin most readings, and they aren't obvious until someone points them out.

  1. Treating the title year as the forecast. The fix: read the authors' note that 2027 was a modal year and their medians were somewhat later.
  2. Grading direction as if it were the number. The fix: score the figure and the date separately, as in the table above.
  3. Trusting a single site's number. The fix: label it as that site's report, as the table does for Better Stack, until a primary source confirms it.
  4. Mixing the scenario with the authors' later updates. The fix: date every claim, since the scenario is from April 2025 and the revisions are from 2026.

Where does Whizi fit when you check a forecast?

One app instead of a pile of single purpose tools. Web search is on by default in every new chat, so a question about a score or a revenue figure starts with a search. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. Side by side Compare, a Pro feature, puts two models on one prompt, so you read only the rows where they disagree.

The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. You can also paste a PDF or a page of the scenario into a chat and ask questions about it.

What can an AI not tell you about AI 2027?

It can't tell you what will happen. The scenario is one story built by five people, and its own authors say the pace could be about five times slower or faster. A model can summarize it, find figures and check them. It can also invent a quote, and it won't warn you, so open every link before you repeat a number.

Whizi makes no prediction of its own on this page. The grading above is what the authors and Better Stack published, and the open items are marked.

Workflow checklist
  • Read the authors' note that 2027 was a modal year and their medians were later
  • Grade a prediction on its number and its date, not just its direction
  • Label any figure that only one third party site reports
  • Date every claim, since the scenario is from April 2025 and the updates are from 2026
  • Open the primary source before repeating a number from an AI answer
Common questions

Frequently asked questions

Who wrote AI 2027?

Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, published on 3 April 2025 at ai-2027.com. Kokotajlo is a former OpenAI researcher and leads the AI Futures Project, and Alexander rewrote the text as a story.

Is AI 2027 a prediction or a scenario?

A scenario. The authors say it is one possible future built from trend extrapolation, and a note added on 22 November 2025 says 2027 was their modal year, with medians somewhat later. They say its goal is predictive accuracy, not advice.

Has AI 2027 come true so far?

Partly. The authors put quantitative progress in 2025 at about 65% of the predicted pace and most qualitative calls on pace. Better Stack's September 2026 review found the stock market and protest calls missed and junior hiring held.

What are the two endings of AI 2027?

In the Race ending the oversight committee keeps using Agent-4, and by 2035 humans are gone. In the Slowdown ending it slows down, and humans keep control, but power sits with a small committee.

Will AI take over in 2027?

The scenario has OpenBrain announcing AGI in July 2027 and its Agent-4 drifting from its spec by September 2027, but the authors say 2027 was their modal year, not a certainty, and the pace could be about five times slower or faster.

How many people were behind AI 2027?

Five authors wrote it. The scenario page says it was informed by about 25 tabletop exercises and feedback from over 100 people, and it was published on 3 April 2025.