AI Judges Changed Their Verdicts on the Same Research Ideas

The research ideas did not change. The AI judge’s verdict often did. In an Ai2-led study, changing the instructions given to GPT-5.4 caused it to return different verdicts on more than half of 154 pairs of human and AI-generated research proposals.

That matters because generating plausible ideas is becoming easier than deciding which ones deserve a scientist’s time. An AI assistant can produce a polished proposal in seconds. If another AI is used to decide whether the proposal is original, its judgment needs scrutiny of its own.

The research, highlighted by Ai2 on October 27, offers a warning about that filtering step. It does not establish that AI cannot invent. It shows how much an apparent measure of invention can depend on the way the evaluator is asked to judge.

The Ideas Stayed the Same. The Instructions Did Not.

The researchers compared two instructions for the same model. One framed the task around expert review: reviewers at a leading AI conference had found one idea novel and the other not, and the model had to identify which was which. The other removed that review and conference framing and simply asked which idea was novel.

Those are meaningfully different instructions, not a punctuation trick. The first gives the judge a prior expectation about what expert reviewers would find. In the human-versus-generated comparison, that framing was part of the experiment; it was not a record of conference reviewers having assessed every generated proposal.

With the reviewer-framed instruction, GPT-5.4 agreed with the benchmark’s expected choice on 93.5% of pairs. With the plainer instruction, agreement fell to 40.9%. The model, ideas and high-reasoning setting stayed the same, and neither condition used a literature search. Each pair was judged repeatedly in both orders.

That is a 52.6-percentage-point drop in the benchmark score. It is not the same quantity as the share of verdicts that changed: the paper reports that the model returned different verdicts on 55.8% of pairs. The distinction matters. A score describes agreement with the benchmark’s labels; a switched verdict shows instability even before deciding whether those labels are right.

The effect was also much smaller when both proposals came from human researchers. On that comparison, the same instruction change moved the score from 83.1% to 76.0%. The especially large swing was not a universal property of every comparison in the study.

This Was Not a Fair Contest Between Human and AI Creativity

The latest preprint, revised on October 26, built its human examples from ICLR 2026 submissions. The higher-novelty pool came from highly rated accepted papers whose reviews contained predominantly positive novelty judgments and no opposing novelty judgment. A contrasting pool came from low-rated rejected papers with the opposite pattern. The judges assessed abstracts, not complete research projects.

For the human-versus-generated test, the researchers paired 154 higher-novelty human ideas with 154 proposals generated by Claude Sonnet 4.5. The model received only a research area and produced an idea in a single pass. It had no literature search, tools or feedback. These were not the best proposals an extensively equipped AI research system could produce.

The benchmark treated those generated ideas as less novel. That is a deliberately weak assumption, not an individually verified fact about every proposal. A blind check by one computer-science professor favored the human idea in all 30 selected pairs, but it did not validate the full set or establish agreement among multiple experts.

These limits make “AI fails to recognize creativity” too sweeping a conclusion. They do not make the changing verdicts disappear. In a separate comparison of proposals from two AI generators, where there was no designated correct winner, changing the instructions also frequently changed which proposal the judges preferred.

Finding Prior Papers Is Not the Same as Understanding What Is New

Giving the judges relevant literature did not reliably solve the problem. In one part of the study, the researchers supplied five earlier abstracts for each idea. Their examination of selected failures found judges treating familiar ingredients as evidence that a new combination was unoriginal. Other judgments mistook the absence of a matching idea in those few papers for evidence of novelty.

Neither mistake means literature searches are pointless. It means a limited search is evidence to interpret, not a certificate of originality. A proposal can combine known methods in a genuinely useful way. It can also sound unfamiliar because the search failed to find its precedents.

This is a different problem from a benchmark changing which tasks it tests. Here, the proposals can stay fixed while the evaluator’s instructions change the result. The measuring instrument is part of the finding.

A Promising Idea Still Has Several Tests to Pass

An AI-generated proposal can be plausible without being new, new without being correct, and correct without being useful. Those are separate questions. The same distinction matters when AI produces mathematical research: generating a convincing argument and establishing a meaningful contribution are not interchangeable achievements.

The practical lesson is to look beyond a novelty score. What did the judge see? Was it told to expect a human-reviewed winner? Which prior work was available? Would a different reasonable instruction change the recommendation?

An AI judge can help researchers sort ideas. But when the same proposal wins or loses because the judging instructions changed, the verdict is a lead to investigate, not a reason to call the idea a discovery.