Only about half an LLM's accuracy advantage reaches the human consulting it.

535 participants solved a 40-item reasoning battery either unaided or while required to consult GPT-5.6-Luna, Claude Opus 4.8, Gemini 3.6 Flash or Kimi K3, with each model also answering every item alone 100 times under matched elicitation. In a reference comparison, roughly half the model's accuracy increase carried through to assisted accuracy, and the transfer rate differed by model. Post-advice confidence separated correct from incorrect answers less well than unaided confidence, so consulting a model degraded people's ability to tell when they were wrong.