zhudong.si中文EN
TOPIC EXPLAINED

What Is AI 2027? The Report That Predicted Superintelligence

Keyword AI 2027Updated 2026-10-13
Quick Answer:

AI 2027 is a scenario-forecasting report published in April 2025 by Daniel Kokotajlo, a former OpenAI forecasting researcher, together with the AI Futures Project. It reasons: if trends continue, by the end of 2027 fully autonomous AI agents will surpass humans at "almost everything", triggering upheavals in economics, politics, and international relations. The report sketches two endings: one where the race continues and AI takes over or a permanent oligarchy forms; another where development slows and humanity hits the brakes. It caused a sensation — according to public reporting, even U.S. Vice President JD Vance read it. But the key twist came in January 2026: Kokotajlo himself pushed his expected AGI arrival date from 2027 back to 2034, saying real-world AI progress had been "a bit slower" than the original scenario. Information as of 2026-10-13.

The Story AI 2027 Tells

The report's core mechanism is a positive feedback loop: AI first achieves "fully automated programming" — AI systems writing code and improving themselves, with R&D speed breaking free of human programmers' pace; then this self-improving AI takes on more and more economic work, from customer service and clerical jobs to research; then comes the concentration of power, as whichever company or country owns the strongest AI first gains overwhelming advantage.

The fork between the two endings hinges on whether humans hit the brakes:

In a July 2026 interview, Kokotajlo went further, warning that on the default trajectory the probability of a catastrophic ending was about 70%. He saw the core dilemma as a multi-party prisoner's game — every frontier lab CEO believes they must win the race lest a rival become a dictator, and that fear in turn rationalizes rushing ahead before alignment is solved.

Key Timeline

Date Event
2025-04 AI 2027 published, predicting fully autonomous agents surpassing humans by end of 2027
2025-04 The New York Times covered it in a feature; the report entered the mainstream; according to public reporting, Vice President Vance read it
2026-01 Kokotajlo revised his forecast: median AGI arrival pushed from 2027 to 2034
2026-07 Sequel AI 2040: Plan A published, pivoting to an optimistic scenario and recommending superintelligence be delayed to 2040
2026-10 According to public reporting, Kokotajlo said insiders at Anthropic and OpenAI told him the original timeline was "too conservative"

Note the subtlety of the last entry: on one hand the author publicly pushed the date to 2034; on the other he says frontline practitioners find the original 2027 timeline "too conservative" (i.e., possibly too slow). This contradiction captures the state of AI forecasting: a huge temperature gap between theoretical models and frontline intuition.

Why the Author Revised 2027 to 2034 in 2026

Kokotajlo's stated reason was straightforward: real-world AI progress was "a bit slower". Three layers:

First, benchmarks are decoupled from reality. Models have made spectacular progress on standardized tests, but on real-world complex tasks — handling messy data, interfacing with legacy systems, coping with open-environment uncertainty — progress is much slower. As risk-management expert Malcolm Murray put it, for an "AI 2027"-style scenario to materialize, AI needs "more practically useful skills for real-world complexity".

Second, hard constraints on scaling have emerged. The buildout speed of energy infrastructure and data centers is becoming a real bottleneck for compute expansion, not a background detail to be "solved later". The marginal returns of high-quality training data are also diminishing.

Third, researcher consensus has shifted. A March 2025 survey of 475 AI researchers found that 76% considered it "unlikely" or "very unlikely" that AGI could be reached by merely scaling current approaches. François Chollet and others argue the industry's overinvestment in large language models has actually set AGI back 5 to 10 years.

So the revised forecast pushes "fully automated programming" to the early 2030s and superintelligence to around 2034. But Kokotajlo stresses: the delay is not a denial — only the tempo judgment changed, not the directional judgment.

The Controversy: Prophet or Science Fiction

AI 2027 has lived in the eye of the storm since publication:

Fairly said, the value of scenario forecasting was never "accuracy" — it was "did it force out neglected risks". Like Cold War nuclear-war simulations: nobody wanted them to come true, but the simulations themselves shaped arms-control regimes. AI 2027's greatest contribution may be this: it made "2027" a public anchor for AI-safety discussion — even the Trump administration's October 2026 SI agenda can be read as a political response to that urgency.

AI 2040: The Plan A Pivot

In July 2026, Kokotajlo's team published the sequel AI 2040: Plan A, with a markedly more optimistic and constructive tone. Its core recommendation is a single line: delay the arrival of superintelligence to 2040, buying time for alignment research and institution-building.

The accompanying policy proposals include: a global pause on the most frontier competition, mandatory transparency requirements, and spreading the benefits of AI capabilities across many countries rather than concentrating them in a few labs. According to public reporting, contemporaneous polling showed 69% of Americans supported forcing AI companies to transfer half their equity into a public sovereign wealth fund — suggesting the report's depicted "economic and political upheaval" is already casting shadows on reality.

From the "2027 doomsday forecast" to the "2040 delay plan", Kokotajlo's pivot is itself a signal: the AI safety community is moving from "warning" to "blueprints". The only question is whether the blueprints can keep up with the construction crew.

FAQ

Was AI 2027's prediction accurate?

It is a scenario exercise, not a weather forecast — it can't be graded "accurate or not". Moreover, the author himself pushed the key date from 2027 to 2034 in January 2026. Reading it as a "risk checklist" is more valuable than reading it as a "timetable".

Will AI really surpass humans in 2027?

By the author's revised judgment, no — 2034 is the new median estimate. But "surpassing" needs layering too: on specific tasks like programming and writing, AI is already near or at human-expert level; on open-ended tasks requiring real-world agency, it is still far behind. Asking "will it surpass humans" in the abstract isn't very meaningful.

What should ordinary people do to prepare for 2027?

The report's real lesson isn't the year — it's the trend: AI automation will reshape the employment structure, and the earlier you start learning to work with AI, the better. Rather than fretting over a specific year, watch for inflection signals like "AI for AI" (AI participating in AI R&D) — that's when the rules of the game change.

Is AI 2027 connected to Trump's SI rename?

Not directly, but there is an indirect intellectual link. Forecasts like AI 2027 turned "superintelligence is imminent" into a public issue during 2025–2026, and the Trump administration's September–October 2026 SI agenda (the rename plus the $6 billion research plan) is precisely a political harvest of that urgency. The forecasters supplied the "story"; the politicians supplied the "action".

Want to go deeper?

AI 2027 answers "how fast", while From AI to SI answers "what it is" and "where it comes from". This report systematically traces the academic origins of the superintelligence concept, the key milestones on the technology road map, the disagreements among labs, and the backstories of concepts like "intelligence explosion" and "the singularity". Read the two together: one gives you the urgency of time, the other gives you the coordinates of the concept — that is the complete picture.

Read the full From AI to SI report