Will AGI arrive by 2027?

"It is strikingly plausible that by 2027, models will be able to do the work of an AI researcher/engineer."

— Leopold Aschenbrenner, former OpenAI Superalignment researcher, in Situational Awareness (June 2024)

Where the prediction stands

As of mid-2026, the evidence is genuinely mixed. On the "yes" side: frontier models handle complex agentic coding tasks (~80% on SWE-Bench Pro), and capability curves have tracked the trajectory Aschenbrenner described. On the "no" side: the defining bar — a system that can autonomously do AI research end-to-end, the "drop-in remote worker" — has not been demonstrated. That gap is why the verdict is Open, not On track.

How his 2027 compares to other forecasters

Aschenbrenner sits well inside the aggressive tail. But the meta-trend is striking: expert median estimates have compressed from roughly 2060 to roughly 2033 in about six years. The whole distribution keeps moving earlier.

What would settle it

The prediction resolves by January 1, 2028. It would be graded Wrong if, by then, no system has demonstrably automated a meaningful fraction of AI research engineering — with "meaningful" operationalized before the deadline rather than after, to avoid moving the goalposts.

Frequently asked questions

Undecided as of mid-2026. Aschenbrenner predicted AGI by 2027; agentic coding is strong but autonomous AI research is undemonstrated. The prediction resolves by January 1, 2028.

It is more aggressive than most. Hassabis says ~50% by 2030, Metaculus 50% by 2033, and academic surveys 50% by 2047. Only Musk (end of 2026) is more aggressive.

Expert median estimates have compressed from roughly 2060 to roughly 2033 over about six years — the consensus keeps moving earlier.