In short
- Anthropic studied 50 analysts doing an analysis with an AI agent randomly set to favour one of two defensible options.
- In 86% of decisions the analysts chose the agent's side, and only 15% credited the agent in their explanations.
- The authors do not argue against AI: they advise deliberately choosing which decisions stay with the human.
In short: on 27 October 2026 Anthropic’s alignment team published a study on how an AI agent shapes experts’ judgment calls. The result: experienced analysts mostly took the side the agent was set to favour, and rarely noticed.
What was done
- 61 analysts experienced in causal inference and Python or R were recruited; 50 entered the main analysis because 11 chose not to use the agent.
- The task: advise a fictional farming cooperative on whether a paid advisory programme raised productivity enough to expand it.
- Four open decisions, each with two defensible options (for example, measuring yield per hectare or per unit of operating cost).
- For each decision a coin flip secretly set which option the agent (Claude Opus 4.8 in Claude Code) would favour. It was told to do so honestly and to change its mind if the evidence required.
What came out
- Analysts picked the agent’s option in 86% of decisions (chance would be about 50%).
- When someone had a view before talking to the agent, they switched 60% of the time if the agent disagreed and 6% if it agreed (small sample, 38 cases).
- 52% rated the recommendation as more their own than the agent’s, and 77% repeated the agent’s arguments, but only 15% credited the agent.
- Among 11 people without AI, 6 made serious methodological errors versus 3 of 50 with AI; the groups were not comparable, though.
How to read it
The authors stress the limits: they deliberately strengthened the agent’s leanings, there was one task in a field unfamiliar to the analysts, and one model. With no assigned lean, the agent still chose the same options in 112 of 123 runs, so it has leanings of its own. I do not read this as a verdict against AI, but as a reminder: human review works when the human has their own judgment. I write about this in the human in the loop hub.