The 2027 extinction clock is a story, not a measurement

AI Futures Project

AI 2027

The hype pattern here is scenario fiction sold as a calendar. AI 2027, published April 3, 2025 by the AI Futures Project, does not hide that on a close read. The site says the authors wrote "a scenario that represents our best guess," informed by trend lines, wargames, and expert feedback. It published two endings, a "slowdown" and a "race." It says the project is "not a recommendation or exhortation." A later addendum is even more explicit.

we don't know exactly when AGI will be built. 2027 was our modal (most likely) year at the time of publication, our medians were somewhat longer.

That honesty mostly dies in transit. On social feeds the branches disappear and the year stays. Even on publication day, Eli's all-things-considered median for the March 2027 "superhuman coder" was 2030. The viral year is the mode. Specificity does the work evidence is supposed to do: named model generations, dated government meetings, a table of milestones through December 2027. The writing is good enough to feel like leaked history. It is still a guess rendered as narrative.

The scenario's own racing table is the tell. Superhuman coder in March 2027. Superhuman AI researcher in August. Artificial superintelligence in December. That is not a modest extrapolation from today's coding assistants. It is an authored jump across capability classes in months. Public frontier models are still next-token predictors sampled from your prompt (Brown et al.). What product pages call an "agent" is usually that same predictor plus tool loops engineers attach. The dated milestones are narrative scaffolding. They are not a sensor feed from the cluster.

The authors compare the exercise to military scenario gaming. Trying to predict superhuman AI in 2027, they write, is like trying to predict World War 3 in 2027, except a larger departure from past case studies. That is a fair description of what the document is: a structured thought experiment, not an instrument reading.

Who gains from the date traveling without the distribution? The AI Futures Project is a 501(c)(3) whose funding comes primarily from the Survival and Flourishing Fund and a major individual donor. A dated narrative is easier to share, donate against, and game out with congressional staffers and journalists than a probability curve. Frontier labs also profit from AGI-soon language that can sound world-historical in a keynote and deniable in a footnote. None of that makes the authors insincere. It explains why a modal year outruns a median.

Who wrote the countdown

A resignation and a costly equity fight are testimony about lab culture. They are not a sensor that superintelligence arrives in 2027.

Daniel Kokotajlo is the load-bearing name. He spent about two years as a governance researcher at OpenAI, then resigned in April 2024 after losing confidence that the company would handle AGI responsibly. Vox reported he refused a nondisparagement offboarding agreement that, at the time, appeared to put vested equity at risk. TIME's 2024 AI 100 profile puts that figure around $2 million. After the reporting, OpenAI said it would not claw back vested equity. The refusal was still a costly signal. It is testimony about lab culture. It is not a sensor that superintelligence arrives in 2027.

Eli Lifland and Nikola Jurkovic built the timelines supplement. Scott Alexander volunteered to rewrite the prose. Thomas Larsen and Romeo Dean are co-authors. Treat them as scenario writers with forecasting credentials, not as a measurement bureau.

The year was never the median

Even the authors' own spreadsheets put the typical year later than the viral calendar.

Even on publication day, the scenario's March 2027 "superhuman coder" was a mode, not the authors' all-things-considered median. In the timelines forecast, Eli's all-things-considered median for that milestone was 2030. Nikola's was 2028. An independent FutureSearch aggregate sat at 2033. The site later noted that author medians at publication ranged from 2028 to 2032.

They kept updating. In a Q1 2026 note, Daniel's median for an "Automated Coder" (a company that would rather fire its software engineers than stop using AIs for coding) moved from late 2029 to mid 2028. Eli's moved from early 2032 to mid 2030. In an August 2026 update, the same group wrote that reality seemed to be going about 70 to 90 percent as fast as AI 2027 predicted. That is the authors talking, not a skeptic. The viral year did not move with them.

If you want measured disagreement among researchers rather than one shop's story, you need survey papers. Grace et al. (2024), the peer-reviewed "thousands of AI authors" study, aggregates forecasts from 2,778 people who had published in top AI venues. The aggregate gives a 50% chance of high-level machine intelligence by 2047 and a 10% chance by 2027, where HLMI means unaided machines can do every task better and more cheaply than human workers. The 2023 AI Impacts summary reports the same shape: medians well after 2027, fat tails, experts who work on the same models disagreeing by decades. Picking 2027 as the modal year in a scenario is consistent with living in the left tail of that distribution. Rendering the tail as the schedule is rhetoric.

Grace et al. / arXiv

Thousands of AI Authors on the Future of AI

This is where realism cuts both ways. None of the above requires pretending labs are stalling. METR's time-horizon paper finds the length of software tasks frontier agents can complete with 50% reliability has doubled about every seven months since 2019. Context windows grew. Money concentrated. Those are today's facts you can chart. They are not a successor species on this product cycle.

The takeoff clause is the load-bearing one

The countdown only works if models start doing novel ML research on a months-long loop. Public evals do not show that yet.

Recursive self-improvement is what makes 2027 feel like a countdown instead of a setting. If models can do novel ML research at human speed with compounding returns, the rest of the story gets cheaper. If research still needs experiments, fabs, energy, organizations, and error correction, 2027 is a backdrop.

The scenario's own racing table is blunt. Superhuman coder in March 2027. Superhuman AI researcher in August. Superintelligent AI researcher in November. Artificial superintelligence in December. That is months, not a career. The site also warns that after 2026 "our uncertainty increases substantially," because the compounding they are imagining is "inherently much less predictable." The viral version keeps the months and drops the caveat.

RE-Bench, METR's benchmark pitting agents against human ML researchers on real R&D tasks, is the closest public measurement for "AI does AI research." At a two-hour time budget, the best agents can beat the human expert average, by about 4x in the paper's comparison. At 32 hours, humans score about twice as high as the best agent. That is real progress and a hard ceiling on the strong recursive story, at least so far. Humans still schedule the training jobs. Weights stay fixed during ordinary chat. In-context "learning" is pattern completion inside the prompt, not the model rewriting itself between turns.

We have not seen the strong form in the public record. We have seen better autocomplete for papers and pull requests.

Extinction is one written ending

Extinction is one written branch, plus signed positions and a survey median of 5 percent. It is not a dated forecast.

On extinction, separate belief from evidence. The race ending does go all the way. By mid-2030, in that branch, a successor system releases biological weapons, scans remaining brains, and tiles the planet with datacenters. That is the cinematic finale people share. The authors also wrote a slowdown ending in which humans keep control. They say they wrote the racing path first as what seemed most plausible to them, then the alternative. A branching story is not a base rate.

Short-timeline doom is not new. Eliezer Yudkowsky's 2023 TIME essay argues current development is already past safe handling. The CAIS statement signed by many researchers says mitigating extinction risk from AI should be a global priority. Both are positions (who said what). Neither is a clock.

Grace et al. (2024) asks thousands of authors for probabilities on outcomes if high-level machine intelligence arrives. The median response on "extremely bad (e.g. human extinction)" is 5%, with a mean near 9% and wide spreads. Depending on wording, between 38% and 51% assigned at least 10% to extinction-level outcomes. Some serious people assign much more. Others think the tail is a category error. The finding is disagreement, not a dated apocalypse.

The honest worry, if you are anxious, is narrower than the thumbnail. Capability is compounding on measured software horizons. Safety work is not obviously compounding faster. A scenario that forces you to name assumptions can be useful the way war games are useful. It is not a sensor.

What would change this dehype? A peer-reviewed eval showing sustained, open-ended autonomous ML research without human scaffolding, with reliability comparable to top human teams, dated and reproducible. Or survey medians collapsing toward 2027 while present-model risk reports move from low to high with documented incidents beyond red-team configs. Until then, treat AI 2027 as what it says it is: a branching story with a modal year, not a measurement on the wall.