Contemporary AI lacks the imagination to diverge or negate in science

A large-scale study involving thousands of working scientists finds that contemporary artificial intelligence systems lack the capacity for divergent thinking and negation required for scientific discovery, despite their potential to accelerate research [1]. The evaluation, mounted by researchers who invited authors of 121,640 recent preprints to judge AI-generated follow-up ideas, is the largest scientist-in-the-loop test of its kind [1]. A total of 6,749 scientists returned 25,139 sets of ratings across biology, medicine, chemistry, and the social sciences, assessing the novelty, empirical feasibility, probability of being true, and favorability of adoption of the machine-generated proposals [1][2]. The findings map what AI cannot yet do for science [2]. Three distinct patterns emerged from the data. Non-reasoning large language models collapsed into a narrow "hivemind" of similar ideas, while reasoning models roamed a wider hypothesis space [1]. However, no model class spontaneously proposed null hypotheses, a move the study notes humans make more freely [1][2]. Scientists also rewarded ideas that resembled their own and prioritized probability over novelty, though social scientists tolerated risk more readily than life scientists [1]. Senior social scientists were the harshest critics, and their skepticism was well-earned: the models faltered most in pluralistic fields like the social sciences that demand context-aware interpretation and evolving theories [1][2]. The study also tested the automated evaluators on which the research community currently relies. LLM-as-a-judge frameworks, artificial metrics, and even state-of-the-art models agreed only weakly with expert judgment, and retrieval augmentation and scientist persona prompting yielded only marginal gains [1][2]. A post-trained reward model, Qwen3-14B, captured field taste nuances and outperformed state-of-the-art models by up to 27%, closing the gap to the inter-rater consistency of independent peer reviewers [1][2]. The work underscores a broader dynamic in scientific evaluation. As the United Nations has noted in the context of the Sustainable Development Goals, tracking qualitative indicators and achieving transformative policy changes remain persistent challenges across complex global systems [9]. The study’s authors conclude that for all the projections, today’s scientific AI remains a collaborator whose imagination, outputs, and judgment benefit from human grounding [1][2].