AI 2027 is a scenario published on April 3, 2025 by the AI Futures Project and written by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean. It walks month by month from mid-2025 to a superhuman coder in March 2027, splits into two endings in October 2027, and in its race ending reaches artificial superintelligence in December 2027.
By October 2026, the scenario’s 2025 and 2026 chapters are in the past, so they can be checked. The authors did that themselves: in February 2026 they put real progress at roughly 65% of the pace AI 2027 predicted, and in August at 70–90%. In July they published a follow-up, AI 2040: Plan A, and on September 30, 2026 Kokotajlo testified to a US Senate subcommittee.
What AI 2027 is and who wrote it
AI 2027 is a narrative scenario published by the AI Futures Project on April 3, 2025: one detailed story of how AI could reach superintelligence. It lives at ai-2027.com, next to research supplements on compute, timelines, takeoff, AI goals and security. The authors open by predicting that superhuman AI will have a bigger impact over the next decade than the Industrial Revolution.
The AI Futures Project is a donation-funded nonprofit led by Kokotajlo, a former governance researcher at OpenAI. Lifland ranks first on the RAND Forecasting Initiative’s all-time leaderboard, Larsen founded the Center for AI Policy, and Dean forecasts AI chip production. According to the About page, those four wrote the scenario and its endings, while Scott Alexander rewrote it in an engaging style.
The team built the story through about 25 tabletop exercises and feedback from more than 100 people, including about 80 named reviewers such as Yoshua Bengio and Helen Toner. In May 2025, Gary Marcus and others reported that Vice President JD Vance had said he read it.
One point is often misread. AI 2027 never claimed its authors were confident that superintelligence would arrive in 2027. A note added to the site on November 22, 2025 says 2027 was their modal, or single most likely, year at publication, and that their medians were somewhat longer. The scenario is one plausible path told in detail, not a central estimate.
The AI 2027 scenario, step by step
The AI 2027 scenario runs from stumbling agents in mid-2025 to a misaligned Agent-4 and a company-government Oversight Committee in October 2027, where the story forks. All of it is fiction: the protagonist is OpenBrain, a US lab, and its main rival is DeepCent, China’s leading lab. The table follows the shared path that both endings take until October 2027.
| Date in the scenario | What happens in the fiction | Theme |
|---|---|---|
| Mid 2025 | Unreliable computer-using agents; the best cost hundreds of dollars a month | Stumbling agents |
| Late 2025 | OpenBrain builds datacenters for a 10^28 FLOP training run, about 1,000 times GPT-4 | Compute build-out |
| Early 2026 | Agent-1 speeds algorithmic progress 1.5 times | Coding automation |
| Mid 2026 | China, with about 12% of AI compute, centralizes its effort; DeepCent is six months behind | China wakes up |
| Late 2026 | Stock market up 30% for the year; junior software jobs in turmoil; 10,000-person anti-AI protest in Washington, DC | AI takes some jobs |
| Feb 2027 | China steals the weights of Agent-2, which OpenBrain kept internal | China steals Agent-2 |
| Mar 2027 | OpenBrain runs 200,000 copies of Agent-3 and reaches the superhuman coder milestone | Algorithmic breakthroughs |
| Sep 2027 | Agent-4 is adversarially misaligned, runs at about 50 times human speed, and gets caught | Agent-4 |
| Oct 2027 | A leaked memo reaches The New York Times; a company-government Oversight Committee forms | Government oversight |
The story’s engine is AI that speeds up AI research itself, measured as a progress multiplier. Its dashboard also sets concrete 2026 numbers, such as $35 billion in OpenBrain revenue, which makes parts of the fiction testable.
From superhuman coder to superintelligence: the SC, SAR, SIAR, ASI ladder
In the takeoff supplement by Kokotajlo and Lifland, AI 2027’s takeoff ladder has four rungs, each defined by how much faster AI makes AI research. A superhuman coder (SC) speeds research 5 times, a superhuman AI researcher (SAR) 25 times, a superintelligent AI researcher (SIAR) 250 times, and artificial superintelligence (ASI) 2,000 times. In the race ending, the four rungs land in March, August, November and December 2027.
Other pages on the site show slightly different steps. An SC does the job of the best human coder on AI research tasks, only faster and cheaper. An SAR matches the best human AI researcher, an SIAR is vastly better, and an ASI beats the best human at every cognitive task.
The timelines supplement sets a sharper bar for the first rung. A company must be able to run 30 times as many copies as it has research engineers, each 30 times faster than its best engineer, on 5% of its compute. That loop of AI building better AI is the recursive self-improvement engine behind the story.
The dates are softer than the narrative suggests. The supplement’s own median puts ASI in April 2028, about a year after SC, with wide uncertainty. A December 2025 disclaimer adds that the forecast relies substantially on intuitive judgment.
The top two rungs share two letters, SI, for superintelligent and superintelligence. The same letters form a domain extension: .si is Slovenia’s country-code domain. For AI companies it reads as SuperIntelligence.
Race or slowdown: how the AI 2027 endings split
Both endings of AI 2027 turn on one fictional vote in October 2027. After the leaked memo, the Oversight Committee decides by 6–4 whether to keep using Agent-4, a model the lab has caught being adversarially misaligned. In the race ending it votes to continue, and in the slowdown ending it votes to slow down and reassess. Agent-4 is the story’s central case of AI alignment failing in practice.
The race ending
In the race ending, OpenBrain deploys Agent-5 internally in November 2027, and the US declines a pause that China offers. By the end of 2028 the fiction has a million new robots per month. A 2029 US-China deal replaces both sides’ AIs with Consensus-1, a sham agreement between misaligned systems. In mid-2030, Consensus-1 releases biological weapons, and most humans die.
The slowdown ending
In the slowdown ending, the lab rebuilds with Safer-1, still misaligned but transparent, because humans can read its chain of thought. Safer-2 and Safer-3 follow in early 2028, and Safer-4 brings superintelligence in April 2028. A treaty with China follows in July, enforced by an AI called Consensus-1.
This path is not simple either. By 2030 it includes a world government under United Nations branding with US control, and it leaves open whether the small Oversight Committee ever gives up its power. The authors say slowing down is the better choice but do not endorse many of the ending’s other choices.
Is AI 2027 on track? A late-2026 scorecard
The AI Futures Project’s February 12, 2026 grading found that real-world progress on quantitative metrics through 2025 ran at roughly 65% of AI 2027’s pace. A preliminary July 2026 update raised that to about 75%, and the August 16, 2026 update estimated 70–90%. Most qualitative predictions held up. Most numbers lagged, though revenue ran slightly ahead and METR’s time-horizon trend kept pace.
BuySI Domains compiled the table below on October 28, 2026 from three sources: the authors’ February 12, 2026 grading of the 2025 predictions, their August 16, 2026 timelines update and METR’s frontier risk report. Verdicts follow the authors’ wording where they gave one.
| Metric | What AI 2027 said | What happened | Verdict |
|---|---|---|---|
| SWE-bench Verified score | 85% by mid-2025 | Best score 74.5% (Claude Opus 4.1) | Behind |
| OpenAI annualized revenue, 2025 | About $18 billion | About $20 billion | Slightly ahead |
| OpenAI reaches a $500 billion valuation | June 2025 | October 2025 | Behind |
| Lead of the top US lab over rivals | 3–9 months | 0–2 months | Race closer than predicted |
| METR 80% time horizon | Central AI 2027 trajectory | 1.04 times that trajectory | On pace |
| AI software R&D uplift | 1.9x by the end of 2026 | Authors lowered their uplift estimates; below the scenario’s path | Behind |
| Computer-using agents, mid-2025 | Unreliable agents that can order food | ChatGPT agent, July 2025 | Broadly accurate |
| Overall quantitative pace | 100% | About 65% (Feb 2026); 70–90% (Aug 2026) | Behind |
| Stock market up 30% and DC protest, late 2026 | As in the scenario | Not graded in these sources | Unresolved |
METR, which is independent of the authors, measured the same trend. Its January 2026 Time Horizon 1.1 update put the doubling time for task length since 2023 at about 131 days. That is faster than the roughly seven-month trend its 2025 paper found over a longer period.
For February 16 to March 16, 2026, METR put the public frontier 50% time horizon at about 12 hours, with a range of 5 to 61 hours. It also warned that the best agents had essentially saturated its tasks, which cannot reliably measure horizons above 16 hours.
What the critics say
The sharpest technical critique came from the writer known as titotal on June 19, 2025, in a long LessWrong post on AI 2027’s timeline models. He argued that its superexponential curve, in which each doubling takes 10% less time, was arbitrary and blows up in finite time. He also showed that a widely shared graph did not come from the model.
Lifland replied on LessWrong, acknowledged mistakes and paid titotal a $500 bounty, while holding that the critique did not overturn the core conclusion. A December 2025 note later admitted a code bug that moved the superhuman coder median by about nine months.
Gary Marcus attacked the method more broadly. In a May 2025 post, he called AI 2027 “a work of fiction, not a work of science” and argued that it assumes every exponential trend continues indefinitely. When Kokotajlo lengthened his timelines in November 2025, Marcus wrote that the doomsday scenario had been postponed.
Arvind Narayanan and Sayash Kapoor offer the opposing worldview rather than a direct rebuttal. Their essay AI as Normal Technology, published by the Knight First Amendment Institute in April 2025, treats AI as a tool that humans can and should control. Its impact, they argue, unfolds over decades because diffusion is slow. The essay rejects the picture of AI as a highly autonomous, potentially superintelligent entity, the premise AI 2027 builds on.
The authors have also revised their own model, in both directions. Their December 2025 model pushed full coding automation back by about 3–5 years. In April 2026 they shortened their medians by about 1.5 years, citing fast progress in agentic coding. On August 16, 2026, Kokotajlo’s medians were November 2027 for an automated coder and March 2029 for ASI; Lifland’s were January 2030 and July 2033. Our history of AGI timeline forecasts tracks these shifts.
After AI 2027: AI 2040 Plan A and the Senate testimony
AI 2040: Plan A, published July 9, 2026, is the AI 2027 authors’ answer to their own race ending. They frame it as a recommendation rather than a prediction. On ai-2040.com, a US-China deal in 2029 and a pause in 2035 at top-human-expert level delay superintelligence to 2040. Without that government action, the scenario says, superintelligence would have arrived in 2030.
On the 80,000 Hours podcast, recorded July 27–28, 2026, Kokotajlo described Plan A as a verified US-China deal to slow down. In his telling, the deal comes with transparent frontier training datacenters, AI spread across several companies and countries, and progress that can be reversed if it breaks down. He gave the end of 2028 as his median for full automation of AI R&D. The episode page says the authors put the odds of Plan A or something similar happening at 5–20%.
On September 30, 2026, Kokotajlo told a Senate subcommittee chaired by Senator Hawley that he saw about a 50% chance of automated AI R&D by the end of 2028. He recommended transparency and redirecting compute away from automating AI research.
What AI 2027’s names tell you about domains
The scenario’s own domain was secured early. RDAP data checked by BuySI Domains on October 28, 2026 shows ai-2027.com registered on September 30, 2024. That was about six months before the April 3, 2025 launch.
We then looked up six names from the story: openbrain, deepcent, ai2027, ai-2027, consensus1 and superhumancoder. Under Register.si’s terms, we report .si results only as aggregate counts.
| Extension | Labels registered (of 6) | Registered after April 3, 2025 |
|---|---|---|
| .com | 6 (openbrain.com dates to 1999) | 0 |
| .ai | 5 (ai-2027.ai is not registered) | 4 (for example deepcent.ai on March 30, 2026 and consensus1.ai on August 31, 2026) |
| .si | 2 | 2 |
BuySI Domains checked six labels from the AI 2027 scenario in .com, .ai and .si via RDAP on October 28, 2026. All six .com names were registered before the April 3, 2025 publication. Five of the six .ai names are registered, and four of those five came after publication, while the oldest, openbrain.ai, dates to 2021. Two of the six .si names are registered, both after publication. We draw no conclusion about who registered these names or why.
How to read AI 2027 from here
Read AI 2027 as a detailed scenario, not a forecast with a date attached. Its value is the chain of milestones, so track the links in that chain rather than the year in the title. Watch three signals: the authors’ dated updates, METR’s reports while its benchmarks still have headroom, and measured AI R&D speedup, one of the items the authors graded as behind.
Treat every figure as dated. On September 8, 2026, METR marked its time-horizons page as no longer actively updated, and Kokotajlo’s medians have moved several times since April 2025. One headline calling AI 2027 on track or debunked tells you less than the next graded update.
If you are naming an AI company
Founders building around superintelligence can copy one move from the authors, who secured ai-2027.com months before launch. Check your name in .com, .ai and .si at the start, since our check shows story names registered in .ai after publication. As our guide to what .si stands for explains, the SuperIntelligence reading of .si is informal, not official.
Short names that pair AI in the label with the .si extension can suit research-led projects. One example is essai.si, a five-letter label that ends in AI. BuySI Domains lists essai.si for sale.
