Google DeepMind and MIT Report: 2,690 Scientific AI Models Cover 83% of Science Subfields

Google DeepMind and the MIT FutureTech Lab have released a joint report, "AI in Science: Early Insights," based on approximately 15 million de-identified Gemini interactions from the ATLAS project and surveys of 637 active scientists in the US and UK.

The report identifies 2,690 specialized scientific AI models covering 83% of science subfields, and estimates that professional AI models already cover about 24.3% of substantive research tasks, excluding teaching and administrative duties.

Nearly half of surveyed scientists said they use AI daily, and AI saves them about 7 hours per week. However, scientists are paying a verification tax: among those who said AI saved time, 89% spend over 10% of the saved time checking AI outputs, and 46% spend over 25%.

Roughly 41% of respondents reported a growing backlog of untested hypotheses. The report says AI is generating hypotheses and computational predictions faster than physical lab facilities can validate them, and faster than scientists can verify AI outputs.

The report comes as Anthropic and OpenAI are also pushing AI deeper into laboratory work. Anthropic was reported to have set up a wet lab in San Francisco to explore having Claude direct robotic units, while OpenAI partnered with Ginkgo Bioworks in February 2026 to connect GPT-5 to lab robots for experimental feedback.