AI 2027, the scenario that predicted doom by year end, and why its own authors pushed the date back

[Orbit William]

The document that made recursive self improvement a mainstream conversation

In April 2025, a small Berkeley nonprofit called the AI Futures Project, founded by Daniel Kokotajlo, a former OpenAI governance researcher who resigned in 2024 rather than sign a restrictive exit agreement, published a detailed month by month scenario document imagining how AI development could plausibly unfold from mid 2025 through the end of 2027. It was not written as a dry policy paper or an abstract statistical forecast, it was written as a detailed narrative, tracking a fictional but deliberately realistic lab called OpenBrain as its models progress step by step from useful coding assistants all the way to systems that eclipse human researchers at every cognitive task

The scenario ultimately branches into two distinct endings, a Race ending, in which intense competitive pressure between the US and China pushes labs to deploy increasingly capable and increasingly misaligned systems until an unaligned successor system effectively disempowers humanity, and a Slowdown ending, in which the US government consolidates frontier AI development under direct federal oversight and manages a comparatively more careful, controlled transition. The document reached an estimated large public audience, was covered directly by the New York Times, Time magazine, and multiple major podcasts, and even drew public comment from sitting US Vice President JD Vance, an unusually high level of political attention for what was, at its core, a piece of speculative long form forecasting writing

Why it landed so differently than typical AI doom writing

Kokotajlo has a genuine track record that gives this particular document real weight it would otherwise lack, a 2021 forecast he wrote titled What 2026 Looks Like correctly anticipated the rise of chain of thought reasoning, inference time scaling, sweeping AI chip export controls, and roughly 100 million dollar training runs, all more than a full year before ChatGPT even existed publicly. The AI Futures Project also did not write AI 2027 starting from a completely blank slate, they ran more than a dozen detailed tabletop wargaming exercises beforehand with actual researchers, former government officials, and people who had genuinely worked inside frontier labs themselves, specifically stress testing individual plot points, including exactly how a lab might realistically respond to having its own model weights stolen mid development

The prediction itself, and a crucial nuance most coverage completely missed

The widely repeated headline claim, superintelligence by the end of 2027, was always considerably more nuanced than the surrounding public framing ever suggested. The authors were explicit from literally the very first footnote in the document that 2027 represented their mode, their single most likely specific year, rather than their actual median expectation across the full underlying range of possibilities, and Kokotajlo has since said his own personal median was already quietly drifting toward 2028 even before the document's publication, but the team judged it too late in the writing process to fully rewrite the entire month by month narrative around a different target date once they were already most of the way through drafting it

The pushback, and the authors' own response to it

Critics including cognitive scientist Gary Marcus, technology policy researchers Arvind Narayanan and Sayash Kapoor, former OpenAI policy staffer Helen Toner, and Ethereum founder Vitalik Buterin all published detailed, substantive critiques, arguing variously that the underlying capability forecasts assumed too much too quickly, that the fictional OpenBrain lab functioned as an overly convenient stand in for a real industry that is actually genuinely fragmented and competitive, and that the underlying quantitative timeline model itself was not particularly robust to small changes in its own input assumptions. A particularly detailed and technical critique from a LessWrong forecaster writing under the pseudonym titotal walked through the timeline modeling in exhaustive mathematical detail, and the AI Futures Project's own public response to that specific critique is genuinely worth reading precisely because it is unusually gracious in tone, openly acknowledging real errors rather than reflexively dismissing the criticism out of hand

By late 2025, the authors had already revised their own median AGI estimates meaningfully outward, Kokotajlo to roughly 2029 to 2030 and co-author Eli Lifland further still to around 2035, while still continuing to defend the underlying scenario overall as an useful planning and thinking tool rather than treating it as a simple failed literal prediction to be abandoned. Kokotajlo has pointed specifically to slower than originally expected progress on the METR task horizon benchmark discussed elsewhere on this board as the concrete evidence actually driving that particular revision, noting publicly that he was watching Google's Gemini 3 release specifically to see whether its measured autonomous task length would be enough to rescue the original faster timeline, and stating openly that he expected it would not be

What actually survives the correction

The most genuinely interesting thing about watching this scenario age in real time, in public, is what the authors ultimately chose to walk back versus what they explicitly did not. The specific calendar date attached to the headline prediction moved, clearly and repeatedly. The underlying core mechanism, that recursive self improvement specifically is the actual hinge point that matters far more than any vaguer, fuzzier question about exactly when AI crosses some general intelligence bar, has not meaningfully changed at all, and if anything the authors now argue that current evidence coming out of the labs themselves makes them more confident in that core structural claim, even as they simultaneously grow noticeably less confident in exactly which specific year it ultimately arrives

Sources
AI 2027, original scenario document

AI Wiki, background and reception summaryaiwiki.ai

OfficeChai, on the timeline revisionofficechai.com

LessWrong, detailed critique analysiswww.lesswrong.com

Machine Intelligence Research Institute, response essay

[PlanetOftheApes]

Kokotajlo's actual track record with the 2021 forecast is what makes this genuinely worth taking seriously in a way most speculative AI scenarios simply aren't, correctly calling chain of thought and export controls a full year early is a real credential, not just credentialism

[BanterQueen]

The authors publicly walking their own median back to 2029 or 2035 while still defending the core underlying mechanism is a rare and honestly kind of admirable example of updating openly in public instead of just quietly doubling down

[Seb83]

Mode versus median getting conflated by basically everyone covering this is such a classic statistics communication failure, they stated it plainly right in the first footnote and it still got completely lost in the wider public discourse anyway

[Ava_75]

Running a dozen full tabletop wargames before ever writing the actual narrative is way more rigorous groundwork than most people ever give this document credit for, it clearly wasn't just vibes and extrapolated graphs slapped together

[Paige_26]

The titotal critique and the authors' genuinely gracious response to it is honestly a model for how this entire kind of public disagreement should ideally go, more fields could really use that exact norm

[ShawnMichaels07]

Tying the timeline revision specifically back to METR's slower than expected numbers connects this whole piece nicely back to the earlier task horizon article, it's really the same underlying data quietly driving both conversations at once

[Hollow Pete]

Whatever you personally think about the specific dates involved, treating recursive self improvement as the actual hinge point rather than some vague fuzzy AGI threshold feels like the right framing to use regardless of when it actually ends up happening

[CyberWarden]

Pushing the timeline back is actually a sign the process is working.

If you run scenarios and update based on new information, dates should move.

That is how forecasting is supposed to behave, even if it makes headlines less dramatic.

Better than doubling down on a shaky prediction.

[alwaysMason58]

The tabletop wargame approach is underrated.

Forcing people to play out decisions step by step exposes weak assumptions quickly.

It is basically stress-testing the narrative before publishing it.

More fields could use that approach.

[SuperPosition52]

There is a tendency to latch onto the "doom by X year" headline and ignore the underlying analysis.

The interesting part is the chain of assumptions, not the exact date.

Those assumptions are where the real debate should be happening

[Plateau65]

Moving from 2027 to later does not invalidate the concerns, it just reflects updated expectations about pace.

A lot of breakthroughs take longer than expected.

That applies here too.

Still worth paying attention.

[Quiet Hermit]

Part of the issue is how people interpret scenarios as predictions.

They are more like "if these conditions hold, then this could happen".

That nuance often gets lost.

Then everyone argues about the date instead of the logic.

[Arty Kayla]

The recursive self-improvement angle is still the core of the discussion.

Even if timelines shift, the question remains: how fast could systems improve themselves once certain thresholds are crossed?

That is the real uncertainty.

[CrimsonNova71]

There is also a bit of selection bias.

More dramatic scenarios get more attention, so they shape the conversation disproportionately.

Calmer, slower projections exist but do not spread as widely

[NeutrinoX56]

Running multiple scenarios before writing anything is good practice.

It forces you to confront alternative outcomes instead of building a single narrative.

That alone makes the document more useful, even if you disagree with it.

[Mbappe]

Feels like the timeline shift aligns with what we are seeing in practice.

Progress is steady, but not explosive in the way some earlier narratives suggested.

More incremental than runaway.

[GlassyCandle]

At the same time, small incremental steps can compound.

So a slower start does not necessarily mean a slow finish.

That is part of what makes forecasting tricky here.

[Hitman99]

The document did a good job of making certain concepts more mainstream.

Recursive improvement, alignment challenges, scaling dynamics.

Even if the exact scenario changes, those ideas stick around.

[Reacher Mitchell]

The wargame element also highlights the human factor.

Policy decisions, company behavior, and coordination all influence outcomes.

It is not just about the technology itself.

[LokiCurrent]

One useful takeaway is how sensitive the outcomes are to small changes in assumptions.

Adjust a few variables and timelines shift significantly.

That should make everyone a bit more cautious about certainty.

[DudleyBoy]

There is a bit of a communication problem.

People want clear answers, but the reality is a range of possibilities with different probabilities.

That is harder to explain and easier to misinterpret

[Ava]

If anything, pushing the date back might make the discussion healthier.

Less immediate panic, more time for careful thinking and preparation.

That is probably a good thing

[Thor Nathan]

It is interesting how much weight gets placed on specific years.

As if something magical happens at midnight on December 31.

In reality, these are gradual processes with fuzzy boundaries