Can AI agents produce scientifically novel and important ideas?

Science News reports that Agents4Science 2025, a virtual conference held Oct. 22, 2025, accepted papers from any field of science on one condition: AI agents had to do most of the work, from formulating hypotheses and analyzing data to the first round of peer review, with humans assessing only the top submissions. By the conference's count, 48 of 314 submissions were accepted, and each team had to detail how people and AI collaborated at every stage. Most of the evidence on quality comes from participants. Min Min Fong's team found AI "really great" for "computational acceleration," but it kept citing the wrong date for when San Francisco's towing-fee waiver took effect, and Fong caught the error only by checking the original source. Risa Wechsler, who helped review submissions, found the papers "technically correct" but "neither interesting nor important," and warned that technical skill can "mask poor scientific judgment."

Organizer James Zou notes that most journals and meetings ban AI coauthors and bar reviewers from relying on AI, to avoid hallucinations, and that this makes it hard to learn how good AI is at science. The conference answers by letting agents carry the early stages and placing human judgment at selection. The pattern the excerpt shows is a split: computation and first-pass work are delegated, while error-checking and judgment about which questions matter stay with people. One counterexample complicates it. Silvia Terragni, a machine learning engineer at Upwork, asked ChatGPT to propose paper ideas from her company's problems, and one result was among the three top papers. "I think [AI] can actually come up with novel ideas," she says.

The excerpt sits on the boundary that Where does AI assistance become unreliable in research? describes. Fong's computational help falls on the reliable side and Wechsler's critique on the unreliable side. Terragni's idea fits neither side cleanly, so the excerpt extends the stage boundary rather than confirming it: the same kind of system is credited with novelty in one case and faulted for unimportance in another. The conference also inverts the arrangement in How should AI agents and humans divide research tasks?, where humans keep most final decisions; here agents do the first-round review and humans judge only the top submissions. It is different evidence from Can specialized agents write better scientific papers than single models?: PaperOrchestra's figures come from a comparison its builders ran, while these verdicts come from outside expert reviewers.

The excerpt does not establish much about quality. It is one journalist's account of one event, with no per-paper scoring, no reviewer agreement figures, no error count across the 48 accepted papers, and no comparison with human-written submissions judged under the same process. The strongest positive claim, that AI can generate novel ideas, rests on one top-three paper described through its author's account. What the excerpt supports is narrower: agents can produce work that clears a human selection process at a venue that required disclosure, and the expert who judged it found it technically sound but missing the judgment that makes work matter. At that strength, autonomy in the early stages of research looks checkable mainly by people with domain expertise, so the verification step remains the bottleneck.