The AI Futures Project, led by former OpenAI researcher Daniel Kokotajlo, has updated its forecast regarding the emergence of a Superhuman AI Researcher (SAR). The collective now predicts that SAR will likely arrive at some point in 2027.
Daniel Kokotajlo shifted his median estimate for SAR by approximately 19 months, assigning a 50% probability of its arrival by May 2027. Contributors Eli Lifland and Romeo Dean set their median expectations for August 2027.
The revision follows a report from the nonprofit organization METR, which estimated that Anthropic is advancing AI capabilities at twice its historical pace. The project now uses a 1.7x AI-driven acceleration factor in its modeling.
The group also updated its projections for TED-AI, a benchmark for systems outperforming humans in economically relevant tasks. Collective estimates suggest a 35–65% probability of TED-AI arriving before 2028.
The researchers noted that their analysis is preliminary and relies on the assumption that major developers will not slow their research. They have called for a pause or slowdown in frontier AI development.
The project clarified that no tasks are currently fully automated at the AL5 level. As of August 2026, 26% of evaluated tasks were at the AL4 level, where AI leads but does not perform tasks autonomously.
