Small prompt changes swing AI verdicts on whether a research idea is new, Ai2 says
The Allen Institute for AI (Ai2) tested six language models as judges of whether a research idea is new. In its study, small changes to the setup gave very different verdicts on the same ideas.
One AI model, GPT-5.4, judged the same pairs of research ideas twice, with two differently worded questions. In a test by the Allen Institute for AI (Ai2), its accuracy fell from 93.5% to 40.9%.
The result comes from a study called 'Old Ideas, Novel Problems', in an Ai2 blog post dated 27 October 2026. It looks at "novelty judges": AI models asked to decide whether a research idea is new compared with earlier work.
Ai2 says there is no standard way to build a judge. It also says these judges are often checked only on papers written by people, not on the AI-generated ideas they are meant to assess.
Ai2 tested six judges: GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5 and Claude Opus 4.6. It says small changes to the setup gave very different judgments of the same ideas. The biggest shifts came from the prompt, meaning how the question was worded.
GPT-5.4 was shown pairs made of one human idea and one AI idea. One prompt said reviewers had found exactly one idea in each pair novel and asked which; the other asked which idea was novel.
With the simpler prompt, accuracy fell below chance, which means worse than guessing. The model also changed its verdict on more than half of the same pairs.
On human-written ideas alone, the same prompt change moved that model's accuracy by only about 7 percentage points. Ai2 says this suggests a judge can look stable on human research yet behave very differently on AI-generated ideas.
The picture was not the same for every model: the same change moved Claude Opus 4.6's accuracy by 21 points on the human-only set.
The test draws on reviews of papers submitted to a research conference called ICLR 2026. Ai2 picked out passages where reviewers discussed originality. Its labels for AI-generated ideas are an assumption: Ai2 says they are weak labels, not verified for every idea.
An expert who independently reviewed 30 pairs, mostly ones the strongest judge got wrong, preferred the human idea every time.
A judge given extra reasoning time and live searching of the arXiv paper archive scored 0.78 on Ai2's accuracy measure, where higher is better. Claude Opus 4.6 scored 0.89, and the searching judge cost roughly 10 times more per decision.
Two purpose-built checkers scored lower on the same measure: Idea Novelty Checker at 0.642, and the novelty checker in Sakana AI's AI Scientist at 0.772.
Ai2 published a benchmark (a standard test), data and code alongside the paper. In its blog post, it says AI novelty judging "remains a major open problem".
Sources
- Novelty bench · allenai.org