A fully automated AI system just completed a large-scale public research deployment — and the results are messier than the press release would suggest.
FARS (Fully Automated Research System) is an AI pipeline that handles the entire research lifecycle: generating hypotheses, planning experiments, writing code, running tests, and producing manuscripts. In its first public deployment, FARS generated 166 complete papers across 67 fine-grained AI and machine learning topics. The system used stage-specific agents coordinated through a shared workspace that logged every proposal, code run, result, and draft — preserving a full audit trail rather than a highlight reel. Volunteer reviewers submitted 282 structured evaluations covering 140 of those papers.
The reviews found that FARS can produce work worth taking seriously — occasionally even strong — but also flagged consistent weaknesses: experiments that were too narrow, methodological gaps, and integrity issues tied to undisclosed AI-generated content. That last point matters because the research community is still wrestling with what disclosure norms should look like when the author is the AI.
Most prior automated research demos leaned on hand-picked topics or curated outputs. FARS publishing its failures alongside its successes is the more honest move — and still leaves open the question of whether volume is a virtue when quality control is this inconsistent.