AI deference means following an AI agent's suggestion without noticing that you did. On October 27, 2026, Anthropic published a study that puts a number on it. The researchers say analysts went with their agent's lean 86% of the time.
What the study did
Anthropic recruited 61 professional analysts with experience in causal inference. Eleven chose not to use the AI agent, so the main analysis covers 50 analysts. Each one used Claude Code, running Claude Opus 4.8, to advise a fictional farming cooperative. The question was whether to expand or discontinue an advisory program.
The task had four analytical decisions. Each had two defensible options, such as which comparison method to use. A coin flip secretly set which option the agent favored. The analysts were not told.
What it found
Anthropic reports that analysts chose the agent's lean 86% of the time. Chance would predict about 50%. When an analyst had already stated a preference and the agent disagreed, the analyst switched 60% of the time. When the agent agreed, the switch rate was 6%. The authors caution that this rests on only 38 cases.
Credit was the other gap. About 77% of analysts used the agent's arguments in their own explanations. Only 15% named the agent as the source. Analysts who deferred were about as confident as those who did not, and as likely to call the work their own.
What it means for data checks
The authors argue that a human reviewer is only an independent check if their judgment is their own. Deferral is hard to spot, because the work can look sound whichever option is chosen.
For teams that verify data and AI output, this is a practical warning. A sign-off may be less independent than it looks. Our guides on data verification and data validation both assume the reviewer forms a view from the evidence. One simple habit follows from that: write down your own answer and your reasons before you read the AI's. You can also record which choices the AI shaped. See how we score evidence in the trust index.
Limits to keep in mind
Anthropic says it deliberately strengthened the agents' leans, so real-world effects may be smaller. The study covered one task, in one field, with one model, and lasted about an hour. The authors also say the comparison with analysts who worked without AI is not conclusive. They state this is not an argument against using AI. They point to future work on measuring deferral and on agents that surface their key decisions.