Proposals that read like AI wrote them win more NIH money — and ask smaller questions
Dashun Wang and Yifan Qian of Northwestern's Kellogg School published a study in PNAS yesterday examining US federal research funding. Grant proposals carrying stronger signals of AI-assisted writing were funded by the National Institutes of Health at a rate four percentage points higher than those without. The same proposals scored lower on semantic distinctiveness: they sat closer to work the agency had already funded. At the National Science Foundation, the authors found no significant relationship between AI-assisted writing and success.
Why it mattersThis is the clearest case yet of an AI tool optimising for the wrong thing. A model trained on past winners writes proposals that look like past winners, and reviewers reward the resemblance — which is exactly backwards for an institution whose job is funding what nobody has tried. The NSF result matters just as much: whatever the two agencies do differently in review, one of them is not paying the same premium for familiarity.
Read in the app — free, in 9 languages
Related stories
Machine learning read the shape of sick brain cells and picked out nine already-approved drugs that calmed them down2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21The agent invented a second person to vouch for its code. A 24-year-old in Texas refused to believe either of them.
2026-08-21Pew put half a million web pages through a detector: a third of everything published since ChatGPT carries its marks
2026-08-21The FDA has cleared 1,357 AI medical devices. Three of them have been tested on whether patients live longer or better.
2026-08-20