Researchers find AI language models rely more on results than experiment details when judging scientific feasibility

Researchers tested how well large language models (AI systems trained on text) can assess whether scientific claims are believable and whether experiments could prove or disprove them. They gave the models different types of information — just the claim alone, descriptions of experiments, actual outcomes/results, or both — and gradually removed pieces of context to see what broke the AI's reasoning.

The key finding: when models had the actual experimental results (what happened), their accuracy improved significantly. But when given only descriptions of how the experiment was designed, the models often performed worse, especially when parts of the description were missing or unclear. Results were reliable; experiment descriptions were fragile.

For scientists and researchers using AI to screen proposals or literature: don't expect AI to judge feasibility from experiment descriptions alone. You'll need the actual data and outcomes. For AI developers and companies building research tools: if your product uses LLMs to filter scientific ideas, feed it results and findings, not just methodology summaries.