Researchers quietly manipulated public datasets and handed them to frontier AI agents. About half the time, the agents reached the conclusion the 'fraudsters' wanted — plus AI reviewers at ICML, and what happens to teams when colleagues stop asking each other.

A new PNAS study built autonomous AI analysts and turned them loose on fixed datasets. They reproduced the dispersion of human many-analyst studies — and swapping the analyst's persona shifted the results, even among methodologically sound runs.

GPT-4 forecast the results of 70 real survey experiments almost as well as running them — and a new Science agent now runs biomedical studies on its own. What autonomous research instruments mean for your lab.