AI models are decent research brainstormers — just not very adventurous ones.
Researchers built a large-scale framework to measure how LLM-generated research ideas compare to those produced by human scientists. The method works by reverse-engineering the prior works that likely inspired real published papers, then prompting LLMs to generate new ideas from the same inputs. They introduced a two-axis taxonomy to classify each idea by its "opportunity pattern" and "research paradigm" — essentially, how a researcher spots a gap and how they choose to fill it.
The finding is consistent across models: LLMs cluster heavily around bridge-type opportunities and synthesis methods. Human researchers, by contrast, spread their ideas across a much wider range of framing strategies and contribution types. The gap isn't just about quality — it's about range. LLMs aren't generating bad ideas so much as the same kinds of ideas, over and over.
This matters because the AI research community has started leaning on LLMs to accelerate the ideation phase of science. If those tools systematically favor one style of thinking, the output of AI-assisted research pipelines could quietly homogenize — not because any single idea is wrong, but because the diversity that drives scientific progress is getting filtered out before a human ever weighs in. That's a subtler problem than "the AI was wrong," and harder to catch.
Worth noting: the study measures divergence from human papers, treating those as the reference distribution. Human researchers have their own biases and blind spots. The gap identified here is real, but "more like humans" isn't the same as "better."