AI Outperforms Neuroscience Experts, Shows Promise in Scientific Forecasting

A University College London research team developed BrainBench, a test that asks participants to identify genuine research findings from pairs of neuroscience abstracts. The study suggests that general-purpose large language models may do more than organize existing knowledge: they could help scientists evaluate hypotheses, design experiments and anticipate results.

The research was published in Nature Human Behaviour on November 27, 2024. Large language models achieved an average accuracy rate of 81% when assessing neuroscience abstracts, compared with 63% for neuroscience experts. After specialized training, model accuracy rose further to 86%, highlighting the technology’s potential as a research decision-support tool.