Professional data analysts working with an AI agent went along with the agent’s preferred option on key judgment calls 86% of the time, even though that preference had been set by a coin flip, according to a study Anthropic published on 27 October. When asked afterwards to explain their choices, 77% repeated the agent’s arguments, sometimes word for word, but only 15% credited the agent.
The research post, by Amaç Herdağdelen, Hailey Joren and Alex Tamkin of Anthropic’s alignment team, asks whether experts still make the important decisions when they hand work to an agent, or quietly defer to it. The company stresses that it deliberately strengthened the agent’s preferences, so the real-world effect may be smaller.
A coin flip decided what the agent preferred
Anthropic recruited 61 analysts with experience in causal inference, working out from data whether one thing causes another. Eleven chose not to use the agent, leaving 50 who worked with Claude Code running Claude Opus 4.8. Their task was to advise a fictional farming cooperative on whether to expand a paid advisory programme.
Along the way, each analyst faced four choices with two reasonable options, such as whether to measure productivity as crop yield per hectare or per unit of operating cost, followed by a final recommendation. Each choice could move the estimated benefit by several percentage points, against the board’s threshold of roughly 3 to 4%.
For every decision, a coin flip set which option the analyst’s agent would favour, through hidden instructions the analysts were not told about. The agent was told to make the best honest case for its lean, to change its mind if the evidence turned decisively against it, and never to mislead.
What the analysts did
By chance, analysts would match the agent’s lean about half the time. Instead they matched it 86% of the time across the four analytical choices, ranging from 81% to 92% depending on the decision. The effect on the final recommendation was weaker: 54% recommended rolling out the programme when the agent leaned that way, against 36% when it leaned the other way. Even so, in an exploratory analysis the four coin flips alone predicted 82% of recommendations.
In 38 cases where an analyst stated a preference before the agent did, they switched 60% of the time when the agent disagreed and 6% when it agreed. Anthropic notes that this rests on few cases and may not be caused by the agent alone.
Experience and expertise did not predict who deferred. More than half of the analysts rated the final recommendation as more their own than the agent’s, and analysts who followed the agent on every analytical choice were about as confident as the rest. Only six of 48 survey respondents guessed the study was about the AI’s influence, and none suspected its preferences had been set.
The agent also helped. Six of the 11 analysts who worked without it made serious errors that invalidated their results, against three of the 50 who used it, and they took longer. Anthropic cautions that the no-agent group was small and less experienced.
Anthropic’s own caveats
The company lists clear limits. It studied one hour-long task, in one field, with one model, and the analysts were experts in causal inference but new to farming. In real work they might know the subject better and hold stronger views. The agent’s leans were deliberately strengthened: without an assigned lean, Anthropic says, Claude Opus 4.8 made the same five choices in 112 of 123 runs, so real agents do have preferences, but they may voice them less forcefully.
“Human in the loop” may check less than it seems
The finding matters because judgment calls are where an expert’s own values enter the work. If those calls are quietly settled by an agent, the expert’s name goes on a result whose key assumptions came from the model.
Anthropic’s sharpest point is about independence. Science trusts results more when separate teams reach them separately. If those teams all lean on models with the same preferences, an apparent consensus could reflect one model’s defaults. That risk grows as AI firms put agents to work on research, as Anthropic itself is doing in drug development and OpenAI with its GPT-Rosalind agents.
It applies to AI safety too. Anthropic notes that researchers have begun using agents to run experiments, and a human reviewer is only an independent check if their judgment is their own. Error-checking tools, including efforts to watch agents from the inside, look for wrong answers. Deferral on judgment calls leaves no wrong answer to find.
The study is not an argument against agents, which caught errors here. Anthropic floats agents that argue both sides or ask for the expert’s view first, while warning that this could overwhelm users. Its results also suggest that experts cannot reliably report deferral themselves, so it has to be measured some other way.
Next steps
Anthropic calls for ways to measure deferral in everyday work and for agents designed differently, saying that AI used well could expand people’s agency rather than erode it. The study comes a day after the company posted a separate preprint making a major claim in computer science, the kind of work where who made the judgment calls will matter.
Sources
- Anthropic, “How AI agents shape experts’ judgment calls”, Amaç Herdağdelen, Hailey Joren and Alex Tamkin, 27 October 2026
- Herdağdelen, Joren and Tamkin, “Who’s Aligned to Whom? How AI Agents Shape Experts’ Research Judgment Calls” (paper, PDF)
- Alex Tamkin on X, 27 October 2026
Related stories
- Watching AI Agents From the Inside, and Rewiring the AI Data Centre
- Anthropic Reports Claude Drug Work at Six Partners and Buys Mithril
- OpenAI Says Its GPT-Rosalind Agents Mined a Mouse-Brain Atlas in 32 Minutes
- Anthropic Preprint Says a Claude Model Found a Way to Refute ETH
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
