Google Research has introduced Science One, an experimental framework meant to make papers written by autonomous AI research agents easier to verify. Instead of letting a model produce a manuscript as a single polished document, the system builds what Google calls a Chain-of-Evidence for the claims it makes.
The goal is to address a practical weakness in autonomous research workflows. Current agents can review literature, propose hypotheses, run experiments and draft papers, but mistakes introduced early can be amplified into plausible-looking results or citations. Science One is designed to preserve the steps behind a claim so another system or human reviewer can inspect them.
Google also described CoE Audit, an automated protocol for checking the integrity of those evidence chains. That matters because the surface quality of AI-written research is improving faster than the tools for verifying it.
This is still a research prototype, not a guarantee that autonomous science is reliable. Its value will depend on whether evidence chains are complete enough for reviewers to find unsupported claims before they become part of the literature.