Pagina nu e încă în română, așa că o vezi în engleză.

AI 2027: what the scenario predicted for 2026, and what actually happened

Quick answer

AI 2027, a scenario published on April 3, 2025 by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean, has moved more slowly than written by its authors' own grading: “roughly 65% of the pace” in February 2026, then “about 70-90%” in August. Some 2026 figures ran ahead, such as Anthropic's revenue. On November 10, 2026, after METR estimated that Anthropic's models are improving about twice as fast as before, mostly because of AI, the authors wrote that a superhuman AI researcher will “likely arrive at some point in 2027”.

Publish my scorecard free

Free plan: no credit card, 1 published page.

On this page

AI 2027 put concrete numbers on 2025 and 2026, so it can be checked. This page sets each prediction next to what happened, with the source and its date. The scenario, its authors' later posts, and every source in the tracker were read on November 16, 2026. This page makes no forecast of its own.

What AI 2027 is, and who wrote it

AI 2027 is a scenario published on April 3, 2025 at ai-2027.com. It tells, month by month, how AI might go from unreliable agents in mid-2025 to superintelligence by the end of 2027, then splits into two endings, a “slowdown” and a “race”. The authors call it “our best guess about what that might look like” and add: “AI 2027 is not a recommendation or exhortation. Our goal is predictive accuracy.”

The site introduces its 5 authors this way:

  • Daniel Kokotajlo, “a former OpenAI researcher whose previous AI predictions have held up well”.
  • Eli Lifland, who co-founded AI Digest and “ranks #1 on the RAND Forecasting Initiative all-time leaderboard”.
  • Thomas Larsen, who founded the Center for AI Policy.
  • Romeo Dean, formerly an AI Policy Fellow at the Institute for AI Policy and Strategy.
  • Scott Alexander, a blogger who “volunteered to rewrite our content in an engaging style”.

The scenario was “informed by approximately 25 tabletop exercises and feedback from over 100 people”. Kokotajlo, Lifland, Larsen, and Dean now work at the AI Futures Project, which Kokotajlo calls “a small research nonprofit”.

What people search for as the AI 2027 report or paper is the scenario itself, its PDF, and 5 research supplements on compute, timelines, takeoff, AI goals, and security. The leading company in the story is fictional: “OpenBrain”, with the others “3–9 months behind”. China's leading company is the fictional “DeepCent”, standing in for DeepSeek, Tencent, Alibaba, and others.

2027 was never the authors' median year. A note added to the foreword on November 22, 2025 says “2027 was our modal (most likely) year at the time of publication, our medians were somewhat longer.” Its footnote adds: “Specifically, our medians ranged from 2028 to 2032.”

What AI 2027 predicted for 2025 and 2026

The scenario calls its forecast “through 2026 … substantially more grounded than what follows”, because for those years it is “heavily informed by extrapolating straight lines on compute scaleups, algorithmic improvements, and benchmark performance.” Here is what each section says, in its own words:

  • Mid 2025, “Stumbling Agents”: “The world sees its first glimpse of AI agents.” They are “impressive in theory (and in cherry-picked examples), but in practice unreliable.” Footnotes put them at 65% on the OSWorld test of basic computer tasks and 85% on SWE-bench Verified, a coding test.
  • Late 2025, “The World's Most Expensive AI”: “OpenBrain is building the biggest datacenters the world has ever seen.” Its new model, Agent-1, is “good at many things but great at helping with AI research”, and its training environments “also make it a good hacker.”
  • Early 2026, “Coding Automation”: With AI help, OpenBrain is “making algorithmic progress 50% faster than they would without AI assistants.” It releases Agent-1, which a footnote puts at 80% on OSWorld, 85% on Cybench, and 1.3 on RE-Bench.
  • Mid 2026, “China Wakes Up”: China keeps “about 12% of the world's AI-relevant compute”, DeepCent is “about six months behind the best OpenBrain models”, and China's leader “sets in motion the nationalization of Chinese AI research.”
  • Late 2026, “AI Takes Some Jobs”: OpenBrain releases Agent-1-mini, “a model 10x cheaper than Agent-1”. “The stock market has gone up 30% in 2026,” “The job market for junior software engineers is in turmoil,” and “there is a 10,000 person anti-AI protest in DC.” The Department of Defense “begins scaling up contracting OpenBrain directly.”
  • Key metrics for 2026: the panel under the late-2026 section shows OpenBrain revenue of $35B for the year, $1T of global AI capital spending (the cost of ownership of active compute), and 38 GW of power for AI worldwide.

AI 2027 tracker: each prediction against what happened

Checked November 16, 2026. OpenBrain is fictional, so each row uses the real company its source names. Where AI 2027's number and the source measure different things, the row says so.

What AI 2027 predictedWhat happenedSource and date
Mid 2025: agents score 65% on OSWorld (footnote 9)The best score at the end of August 2025 was 60.8%, on OSWorld-Verified, a corrected version of the same testAI Futures Project grading sheet, February 12, 2026
Mid 2025: “agents will score 85% on SWEBench-Verified” (footnote 10)The authors: “the best actual score was 74.5% (Opus 4.1)”Lifland and Kokotajlo, “Grading AI 2027's 2025 Predictions”, February 12, 2026
Late 2025: other companies “3–9 months behind OpenBrain”The authors: “more like a 0-2 month lead between the top US AGI companies”Grading post, February 12, 2026
Late 2025: Agent-1's training makes it “a good hacker”Anthropic says Claude Mythos Preview can “surpass all but the most skilled humans at finding and exploiting software vulnerabilities”. Anthropic says it doesn't plan to make the model generally availableAnthropic, Project Glasswing, April 7, 2026
Early 2026: Agent-1 is released and scores 80% on OSWorld (footnote 29)Anthropic reports 79.6% on OSWorld-Verified for Mythos Preview, which isn't generally available, and 72.7% for Claude Opus 4.6Anthropic, Project Glasswing, April 7, 2026
Early 2026: AI makes OpenBrain's algorithmic progress 50% fasterMETR: Anthropic is “likely improving the capabilities of their models on well-specified tasks at around twice the rate” of March 2024 to October 2025, and “perhaps ≈80%” of today's acceleration is due to AI. Anthropic “redacted a majority of the evidence” METR found most informative. The two numbers measure different things; the authors read METR's as about 5 months ahead of the scenarioMETR, November 5, 2026; AI Futures Project, November 10, 2026
End of 2026: “~3 work week 80% coding time horizons” on a central AI-2027-speed trajectory of the authors' model (as their grading post states it)METR's last published measurement: Claude Mythos Preview (early), with an 80% time horizon of about 186 minutes. METR says its page “is no longer actively updated”METR time horizons, measurement added May 8, 2026
Mid 2026: China has about 12% of the world's AI-relevant computeEpoch AI: “Since 2022, 72% of AI compute has ended up in the US and about 10% in China, including chips China makes itself”Epoch AI, October 27, 2026
Mid 2026: DeepCent is about 6 months behind OpenBrainEpoch AI: the best Chinese model, the open-weight Kimi K3, “trailed the frontier by only 4.4 months” on the Epoch Capabilities Index. From 2023, the gap averaged 7 months, with a low of 4 and a high of 14Epoch AI, September 23, 2026; Epoch AI, January 2, 2026
2026: OpenBrain revenue of $35B for the year, or “$55B in annualized revenue” by 2026 in the authors' grading postAnthropic's revenue was $33.2 billion in the 9 months to September 30, 2026, and its run-rate revenue was “approximately $81 billion as of October 31, 2026”. OpenAI said on March 31, 2026: “We are now generating $2B in revenue per month”Anthropic's S-1/A, November 10, 2026; OpenAI, March 31, 2026
2026: a “valuation of $2.5T” for the leading company (the authors' grading post)OpenAI raised $122 billion “at a post money valuation of $852 billion”. Anthropic's IPO price range is $70 to $78 a share, set when its roadshow began; the offering hadn't priced thenOpenAI, March 31, 2026; Anthropic, November 10, 2026
Late 2026: “The stock market has gone up 30% in 2026”The S&P 500 closed 2025 at 6,845.50 and closed November 13, 2026 at 7,561.67, a rise of 10.5%FRED, data from S&P Dow Jones Indices, close of November 13, 2026
Late 2026: “The job market for junior software engineers is in turmoil”Stanford: employment of workers aged 22 to 25 in highly AI-exposed jobs “now stands about 19% below where it would be” had it kept pace with less-exposed peers, in data through June 2026. The update's title starts “No Widespread Displacement”Stanford Digital Economy Lab, August 12, 2026
Late 2026: “a 10,000 person anti-AI protest in DC”No report of one found. Decrypt counted “About 200 protesters” at a march in San Francisco in JulyDecrypt, July 13, 2026; searched November 16, 2026
Late 2026: the Department of Defense contracts OpenBrain directly, through an Other Transaction Authority contract (footnote 41)The DoD's Chief Digital and Artificial Intelligence Office awarded Anthropic “a two-year prototype other transaction agreement with a $200 million ceiling”. On March 4, 2026, the Department of War told Anthropic it had been designated a supply chain risk, and Anthropic said it would challenge that in courtAnthropic, July 14, 2025; Anthropic, March 5, 2026

The authors keep their own grading of the 2025 predictions, with sources for each, in the spreadsheet linked from their February post. They hadn't published a grading of the 2026 predictions by November 16, 2026.

Is AI 2027 on track? What its authors say now

The authors have posted an update every few months on the AI Futures Project blog. In date order:

  • July 2025, a note on ai-2027.com: “We've made some updates to the forecast which push the median back 1.5 years while maintaining SC in 2027 as a serious possibility.” SC is their superhuman coder milestone.
  • November 22, 2025, the foreword note: 2027 was the modal year, and their medians were 2028 to 2032.
  • January 27, 2026, a correction of press coverage in The Guardian, The Independent, Inc., The Washington Post, and the Daily Mirror: “it has never been the case that we were confident AGI would arrive in 2027.”
  • February 12, 2026, the grading of 2025: “In aggregate, progress on quantitative metrics is at roughly 65% of the pace that AI 2027 predicted. Most qualitative predictions are on pace.” A July 2026 note on the same post moves that to “more like 75%”.
  • April 2, 2026, the Q1 update: Kokotajlo's median for an automated coder moved “from late 2029 to mid 2028”, and Lifland's “from early 2032 to mid 2030”.
  • July 9, 2026, AI 2040: Plan A, a second scenario in which humanity delays superintelligence until 2040. The authors call it “primarily a recommendation, not a prediction.”
  • August 16, 2026, the Q2.5 update: “reality seems to be going about 70-90% as fast as AI 2027 predicted.”
  • September 30, 2026, Kokotajlo testified to a Senate subcommittee that AI companies are starting to train AIs to do the whole research process: “I personally would guess about 50% chance by end of 2028.”
  • November 10, 2026, the Q3.5 update, titled “Looks Like the Intelligence Explosion Will Happen in 2027 After All”: “all of us forecast that Superhuman AI Researcher (SAR) will likely arrive at some point in 2027 (assuming no substantial deliberate slowdown).” They put a 35-65% chance on Top-Expert-Dominating AI (TED-AI), “AIs that can automate all white-collar work”, before 2028. The METR report was “the single biggest source” of the change.

The November 10 post gives its authors' medians for TED-AI, which they use as their AGI milestone. Brendan Halstead, a co-author of the post who isn't among AI 2027's authors, is one of them:

AuthorTED-AI median, November 10, 2026Earlier median, as the post gives it
Daniel KokotajloSeptember 2027Late 2028
Eli LiflandSeptember 2028Mid-2032
Thomas LarsenApril 2028Early 2032
Romeo DeanSeptember 2028Early 2035
Brendan HalsteadApril 2028Early 2031

The post says these forecasts are “conditional on people not ‘pacing the frontier’ (i.e., delaying when a given capability level is achieved) any more in the future than in the recent past.”

The main criticisms of AI 2027

  • The timelines model. A computational physicist who writes as titotal wrote on June 19, 2025 that “the fundamental structure of their model is highly questionable and at times barely justified, there is very little empirical validation of the model, and there are parts of the code that the write-up of the model straight up misrepresents.” Lifland and Kokotajlo replied on December 16, 2025: “we disagree with most of titotal's criticisms.” The main fix they made because of it “would have pushed back the median Superhuman Coder arrival time by about 9 months.”
  • A story, not a forecast. Gary Marcus wrote on May 22, 2025: “It's a work of fiction, not a work of science.” On November 21, 2025 he wrote that the scenario “has officially been postponed, by its co-creator”. The authors' January 27, 2026 post answers coverage of that kind.
  • AI as a normal technology. Arvind Narayanan and Sayash Kapoor, the authors of “AI as Normal Technology”, wrote with Kokotajlo, Lifland, and Larsen on November 12, 2025 that they all agree “Before strong AGI, AI will be a normal technology.” They part ways after that: Narayanan and Kapoor “think strong AGI won't be developed in the lab (say, by scaling LLMs)”, and they think today's benchmarks “have poor construct validity”.
  • Redacted evidence behind the November 2026 update. METR wrote that “Anthropic redacted a majority of the evidence we found most informative”, so “the public has to rely on METR's judgment much more than would be ideal.” Larsen wrote in the November 10 post that “the METR report on its own does not have sufficient data in it to support that claim (because of redactions).” The claim is SAR before the end of 2027.

AGI timeline forecasts, with dates

These are other people's forecasts, each with its date; this page adds none. Each Metaculus question sets its own tests for “general AI”, so read the question before comparing its date with AI 2027's milestones.

WhoWhat they forecastDate
Metaculus community, “When will the first weakly general AI system be devised, tested, and publicly announced?”August 9, 2027 (25th to 75th percentile: September 10, 2026 to September 5, 2029), from 1.8k forecastersRead November 16, 2026
Metaculus community, “When will the first general AI system be devised, tested, and publicly announced?”June 1, 2030 (25th to 75th percentile: February 5, 2028 to March 2036), from 1.9k forecastersRead November 16, 2026
Daniel Kokotajlo, AI Futures ProjectTED-AI median: September 2027. Superhuman AI researcher median: May 2027November 10, 2026
Eli Lifland, AI Futures ProjectTED-AI median: September 2028. Superhuman AI researcher median: August 2027November 10, 2026
Anthropic“we expect powerful AI systems will emerge in late 2026 or early 2027”March 6, 2025
Dario Amodei, Anthropic's CEONo year. “since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI”September 2026
Sam Altman, OpenAI's CEO“2026 will likely see the arrival of systems that can figure out novel insights. 2027 may see the arrival of robots that can do tasks in the real world.”June 10, 2025
Demis Hassabis, Google DeepMind's CEO“My estimate is sort of 50% chance by in the next five years, so by 2030 let's say.”July 23, 2025, Lex Fridman Podcast #475

Keeping your own scorecard? Ask Claude, ChatGPT, or Gemini to turn your table into one HTML file, then publish it with Convika. When a number changes, Upload as a new version keeps the page at the same link.

Frequently asked questions

What is AI 2027?

AI 2027 is a scenario, published on April 3, 2025 at ai-2027.com, that describes month by month how AI might reach superintelligence by the end of 2027. It comes with a PDF, 2 endings (“slowdown” and “race”), and 5 research supplements on compute, timelines, takeoff, AI goals, and security. Its authors say its goal is “predictive accuracy”, and that 2027 was their most likely year, not their median.

Who wrote AI 2027?

Daniel Kokotajlo, a former OpenAI researcher, wrote it with Eli Lifland, Thomas Larsen, and Romeo Dean, and Scott Alexander rewrote it “in an engaging style”. Kokotajlo leads the AI Futures Project, the nonprofit that publishes the authors' later updates.

Is AI 2027 coming true?

By its authors' own grading, partly and more slowly than written. On February 12, 2026 they put progress on its numbers at “roughly 65% of the pace” and said “Most qualitative predictions are on pace.” In August 2026 they put it at “about 70-90%”. On November 10, 2026, after METR's report on Anthropic, they wrote that AI-driven acceleration is about 5 months ahead of the scenario and that a superhuman AI researcher will “likely arrive at some point in 2027”.

Did the AI 2027 authors update their predictions?

Yes, every few months. Their medians moved later through 2025, then earlier through 2026. On April 2, 2026, Kokotajlo's automated-coder median moved from late 2029 to mid-2028. On November 10, 2026, all of them forecast a superhuman AI researcher in 2027, and Kokotajlo's median for TED-AI, their AGI milestone, became September 2027.

Is there an AGI 2027 tracker?

Yes. This page's table checks 15 of AI 2027's predictions for 2025 and 2026 against dated sources. The authors grade their own 2025 predictions in a public spreadsheet linked from their February 12, 2026 post. The independent AI 2027 Tracker follows 66 predictions and was last updated on November 9, 2026.

When do experts predict AGI?

The dates differ by who forecasts and by how they define AGI. Metaculus's community puts a “weakly general AI” at August 9, 2027 and a “general AI system” at June 1, 2030, read November 16, 2026. The AI 2027 authors' medians for TED-AI run from September 2027 (Kokotajlo) to September 2028 (Lifland and Dean). Demis Hassabis gave a 50% chance by 2030 in July 2025.

What is OpenBrain in AI 2027?

OpenBrain is the fictional AI company at the center of the story, used “to avoid singling out any one existing company”. The scenario imagines its rivals “3–9 months behind”. China's leading company in the story is the fictional DeepCent, which stands in for DeepSeek, Tencent, Alibaba, and others.

Where can I read the AI 2027 PDF?

The PDF is at ai-2027.com/ai-2027.pdf, linked at the top of the scenario. The site also has a Listen version, and a summary and the 5 research supplements under Supplements.

Next step

How to password protect an HTML page or website

A JavaScript password script leaves the password in the page source. To password protect an HTML page or link, encrypt the file or let a server check it.

Related articles

More ways to send your page

Skipbox

Cheap & fast GitHub Actions runners, unlimited for $39 a month

Your own machine keeps Docker layers and package caches between jobs. Switch by changing one runs-on line.

Start a free Skipbox trial⁠

Share this article

Keep your AI 2027 scorecard at one link

Ask your AI to turn your table into one HTML file and upload it. Convika publishes it, and each new version goes live at the same link.

Publish my scorecard free

Free plan: no credit card, 1 published page.