Last month Anthropic reported that 77% of the analysts working with its AI agent repeated the agent's argument in at least one answer, sometimes word for word, and that "only 15% credited the agent anywhere." The figures come from a paper Anthropic published on 27 October 2026, and Al Jazeera reported them the next day.
Read them with the disclosure first. Anthropic ran the study on its own product. It gave the agent a lean on each decision through hidden instructions that told it to hold the lean as its own view, set by coin flip, and the analysts were not told. Anthropic wrote: "we strengthened their preferences on purpose to measure their effect."
The headline number is Anthropic's: "Across the four analytical choices, analysts went with their agent’s lean 86% of the time." If the agent made no difference, Anthropic says, about 50% would be expected. Fifty analysts used the agent, from 61 recruited, and the task was to advise a farming cooperative on whether an advisory service was effective and should be expanded.
Kant, in "What is Enlightenment?" (1784), supplies the mechanism. "Enlightenment is man's release from his self-incurred tutelage." Tutelage, he wrote, is the inability to use one's understanding without direction from another, and his examples are plain: "If I have a book which understands for me, a pastor who has a conscience for me," then "I need not trouble myself." He placed the fault in a lack of resolution and courage, not of reason. Our reading is that an analyst who reproduces the agent's argument without saying whose it is has let another understanding stand in for their own. Kant wrote nothing about AI. The mechanism is what carries over, and the Beck translation on Wikisource is where to check it.
The sharpest figure is also the weakest. In 38 cases an analyst had stated a preference before the agent did. When the agent's lean went against that choice, Anthropic reported, they switched 60% of the time, against 6% when it agreed. Anthropic's own caution: "because this rests on only 38 cases, the estimate is imprecise."
The limits are the authors' own. "We studied one task, in one field, with one model." Estimates are averages across analysts, and Anthropic recorded which option analysts chose but not how strongly they preferred it. Because the lean was strengthened on purpose, the 86% measures what a reliably opinionated agent can do, not what agents do in ordinary work. Our reading adds a further limit: choosing the agent's option may sometimes be persuasion by a good argument, which the design does not separate from deference. Only six of 48 analysts guessed the study concerned the AI's influence on their choices, and none suspected its preferences had been manipulated. Anthropic's researchers, as reported by Al Jazeera, said such deference may be "hard to identify, even for the experts themselves."
For the wider question of when a recommendation bypasses a chooser's reasons, see our glossary entry on algorithmic nudging and our comparison of hard determinism and compatibilism on algorithmic choice.
Our position: a human in the loop is a check only while the judgement is the human's own, and this study shows how far that can slip in one hour-long task, not how far it does in practice. Watch for: a replication by someone other than the vendor, with an agent whose lean the vendor did not set, and with credit to the agent measured as well as choices.
Cover image: Man at a laptop in an office, Bench Accounting, CC0, via Wikimedia Commons.
.jpg?ref=philocrux.com){kind=link}