Cosmos Institute Challenges AI-Run Science

Verification is a necessary gate, not a substitute for scientific judgment. The essay identifies a failure mode benchmark culture often hides: a correct result can remain unusable if nobody can explain, defend, or responsibly apply it.

• –AI scientists need persistent identity and memory so claims can be challenged and retracted over time.
• –Human review is part of knowledge formation, not merely a safety brake.
• –AI evaluations should measure inspectability, assumption disclosure, and revision behavior alongside accuracy.
• –Autonomous discovery could optimize for verifiable outputs while narrowing the questions scientists ask.
• –Developers should treat accountability and interpretability as core system features.