Google Cloud Unveils Science One to Cut AI Paper Hallucinations From 21% to 0%
Updated
Updated · Google Research · Jul 30
Google Cloud Unveils Science One to Cut AI Paper Hallucinations From 21% to 0%
2 articles · Updated · Google Research · Jul 30
Summary
Google Cloud introduced Science One and the Chain-of-Evidence Audit, an experimental framework that ties every AI-generated research claim to recorded evidence instead of adding citations and support after writing.
Tests on 75 papers across five systems-optimization tasks found baseline agents hallucinated up to 21% of references, while Science One produced zero phantom citations, fully verifiable scores and the strongest method-code alignment.
Science One achieves that by combining a literature-grounding module, a discovery engine and a paper-writing verifier that checks claims against source papers, code and experiment logs.
Google said the stricter verification did not hurt performance: the prototype matched or beat human experts on all five ADRS tasks and posted the best overall score on Cloudcast and EPLB.
The company framed Science One as a research prototype rather than a product, arguing verifiability will become a core design requirement as autonomous AI systems take on more end-to-end scientific work.