Co-Scientist cuts severe hallucinations to 4%

Google DeepMind’s Co-Scientist cut severe result hallucinations to 4% across 150 generated manuscripts, versus 90% for the Agent Laboratory baseline, in a double-blind study with 30 experts and 450 reviews. The paper says its safety system also refused 98.7% of harmful research directions; builders should treat execution logs and human review as core controls, alongside automated alignment research’s $4-per-hour benchmark.