A classic creativity trick - borrow an idea from something totally unrelated - works differently on chatbots than it does on people, according to a new study.
Researchers had human participants and seven large language models design products two ways: either drawing inspiration from a random, unrelated source or simply addressing an unmet user need. Humans came up with more original ideas whenever they had to make that random cross-domain leap, and the boost held at nearly any distance between source and target. LLMs were more original than humans overall, but the random-source prompt gave them no consistent advantage. It only helped once the source idea was semantically distant enough, and even then only the highest-rated models capitalized on it. The two groups also borrowed differently: humans lifted surface features, like what something looks like, while LLMs pulled structural and functional properties, like how something works.
This matters because "add a random constraint" is standard advice for prompting chatbots toward more original output, on the assumption that a trick from human psychology transfers directly to machines. This study suggests it doesn't, and that raw model capability - not just clever prompting - determines whether an LLM can exploit a distant analogy at all.
Weaker models, in other words, may simply lack the reach to turn a strange comparison into a usable idea, no matter how the prompt is worded.