No. 030

AI capability in research tasks is advancing faster than institutional frameworks for governing it. Google released Deep Review, a public tool that checks manuscripts for mathematical errors, logical gaps and unsupported claims, built on a system already piloted on more than 15,000 conference submissions. TasteVal, a new benchmark for experimental research design, finds frontier models now exceed the human expert baseline, with scores tripling in less than a year. Separately, 29% of US STEM PhD dissertations filed in 2026 contain AI-generated writing, up from near zero before 2023, and graduates whose dissertations include AI writing are less likely to enter academic research careers. These findings arrive the same week that a national faculty survey found researchers adopting AI faster than their institutions are providing guidance on its use.

Tools and Capability

• Google launches Deep Review AI tool to help researchers catch manuscript errors

  R&D World, October 14 2026

  Google Research released Deep Review, which checks technical documents for mathematical errors, logical gaps and unsupported claims, with two free reviews per week. Its predecessor, piloted at STOC, ICML, NeurIPS and ICLR, found 89.7% of known proof errors on a small benchmark (29 errors in 26 papers) against 55.2% for a single Gemini prompt, and pilot users reported cases where it wrongly challenged correct proofs.

• TasteVal: Measuring the Experimental Research Taste of AI Systems Against Human Experts

  arXiv, October 5 2026

  A benchmark pitting frontier AI models against 24 human domain experts on open-ended experiment design finds models now exceeding the human expert baseline, with scores roughly tripling in less than a year, suggesting AI is approaching the capacity to direct autonomous research programs rather than merely assist them.

Integrity and Quality

• PhD students who write dissertation with AI less likely to stay in academia

  Nature, October 16 2026

  A study of 226,000 US STEM dissertations finds AI-generated writing appears in 29% of those filed in 2026, up from near zero before 2023, and graduates whose dissertations include AI writing are 17% less likely to enter academic research careers and 16% more likely to enter industry.

Institutional Response

• Supporting researchers in a time of change

  Ithaka S+R, October 16 2026

  A national survey of 4,003 US researchers finds AI adoption is rapid but institutional governance lags behind, with the clearest need being coordination across faculty, libraries, IT, and research integrity staff rather than any single policy intervention.

• AI in the workplace vs. AI in education

  Paired Ends, October 11 2026

  Drawing on cognitive science research, Turner distinguishes the workplace case for AI (where the outcome is what matters) from the educational case (where the process builds the judgment that future research requires), an argument with direct bearing on how researchers at different career stages should think about what tasks to delegate.