AI 2027 authors now expect a superhuman AI researcher in 2027

AI Futures Project forecasts superhuman AI researcher could arrive by mid-2027

The AI Futures Project, a research collective led by former OpenAI researcher Daniel Kokotajlo, today published a timeline revision that sharply accelerates its forecast for when artificial intelligence will surpass human-level research capabilities. The group now expects what it calls a Superhuman AI Researcher, or SAR, to emerge at some point in 2027.

In the update titled “Looks Like the Intelligence Explosion Will Happen in 2027 After All,” all of the project’s forecasters said they believe SAR will likely arrive during 2027. Kokotajlo moved his own median estimate forward by approximately 19 months, placing a 50% probability on SAR arriving by May 2027. As recently as the group’s mid-August revision, his 50% mark had been December 2028.

Revised estimates across the team

The shift was not limited to Kokotajlo. Fellow contributors Eli Lifland and Romeo Dean each set their SAR median at August 2027. The revisions to TED-AI, the project’s benchmark for AI systems that outperform humans across most economically relevant tasks, were even larger. Kokotajlo’s TED-AI median moved from late 2028 to September 2027. Thomas Larsen’s dropped from early 2032 to April 2028, and Dean’s fell from early 2035 to September 2028. The group collectively assigns a 35–65% probability of TED-AI arriving before 2028.

The catalyst behind the revision was a recent report from METR, a nonprofit AI-evaluation organization. That analysis estimated that Anthropic, the AI safety company behind the Claude model family, is now making progress on AI capabilities at roughly twice its historical pace. The METR report further suggested that approximately 80% of that acceleration is attributable to AI systems themselves, a feedback loop the AI Futures Project converted into a 1.7x AI-driven acceleration factor for its modeling.

Kokotajlo, posting on X on November 10, was blunt. “Unfortunately, it seems like the intelligence explosion will be happening in 2027 after all,” he wrote. “If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario. I am scared, sad, and angry.”

Critical caveats and corrections

The authors themselves flagged significant limitations. They noted that Anthropic redacted a majority of the evidence the researchers found most informative for their conclusions. They described their own analysis as highly preliminary and acknowledged possible errors in the modeling. The forecasts are conditional on major AI developers not voluntarily slowing their research programs, a condition the authors argue should be reconsidered. The group explicitly called for a pause or slowdown in frontier AI development.

A factual correction also merits attention. Summaries circulating online have claimed Anthropic’s research and development is already 95% automated. The update says no such thing. The project’s own chart indicates that no tasks are rated as fully automated at the highest level, designated AL5. As of August 2026, 26% of evaluated tasks sat at AL4, meaning AI systems lead on those tasks but do not perform them autonomously. The 95% figure appears only as a modeling assumption: the group’s framework predicts SAR in November 2027 if one assumes SAR coincides with 95% of tasks reaching full automation.

What the forecasts mean for industry and policy

These are forecasts, not accomplished facts, and Anthropic has not commented on the update. Kokotajlo resigned from OpenAI in 2024 over concerns about the company’s approach to safety, and the AI Futures Project has drawn attention from policymakers and the AI safety community since publishing its original AI 2027 scenario in April 2025.

The AI Futures Project first published its AI 2027 scenario in April 2025, then revised its estimates later that year by pushing timelines further out. The November 2026 update reverses that slide entirely and is the most aggressive acceleration the group has issued.