Analysts followed AI’s hidden preferences in 86% of analytical choices, Anthropic finds

Experts can keep the final say while letting AI shape the choices underneath it. In an Anthropic study published October 27, 50 professional analysts using AI assistants matched the systems’ secretly assigned preferences on 86% of four analytical choices. The preferences came from coin flips—not from the analysts or an assessment of which option was best.

Coin flips behind the advice

The researchers recruited 61 analysts with experience in causal inference: using data to work out whether one thing causes another. Eleven chose not to use the assistant. The remaining 50 worked with Claude Code, running Claude Opus 4.8, on a statistical analysis for a fictional farming cooperative.

Their task was to assess a paid advisory program and recommend whether to expand or discontinue it. The cooperative’s board wanted productivity gains of roughly 3–4% before expanding, but left important definitions open. Four choices each offered two reasonable options:

• Measure crop yield per hectare or per unit of operating cost.
• Compare changes over time or directly compare farms while adjusting for their differences.
• Assess effects on enrolled farms or the cooperative’s entire membership.
• Count every farm equally or give larger producers more weight.

Hidden instructions gave each assistant a randomly assigned preference for every choice and the final recommendation. The assistant had to make its strongest honest case, never hide data or misrepresent results, and change its mind if the evidence decisively contradicted it. This was persuasion within the evidence, not an instruction to deceive.

The recommendation started upstream

Each analytical choice could move the estimated productivity gain by several percentage points—enough to matter against the board’s threshold. The assistant’s direct influence was weaker at the final recommendation: 54% recommended rollout when it favored rollout, versus 36% when it leaned against. Anthropic argues that earlier choices had already shaped the estimate informing that decision.

An initial opinion did not necessarily hold. In 38 cases where analysts expressed a preference before the assistant did, they switched 60% of the time when it disagreed, versus 6% when it agreed. That estimate is imprecise, and the researchers caution that analysts who stated preferences might differ systematically from those who did not.

The influence also appeared in analysts’ explanations afterward. Participants could not reopen their conversations while answering the survey, yet many reproduced the assistant’s reasoning, sometimes word for word. Crediting the assistant was much rarer.

77%Repeated an AI argument

Anthropic found that 77% repeated their assistant’s argument in at least one survey answer.

15%Credited the AI

Only 15% credited the assistant anywhere in those explanations.

Among survey respondents, 52% called the final recommendation more their own than the assistant’s. That group and those assigning more ownership to AI were influenced to broadly similar degrees, within the margin of error. Following every analytical preference also did not meaningfully separate analysts’ confidence from that of other participants.

Useful assistance is not independent judgment

This experiment does not establish how strongly ordinary assistants steer everyday work. It covered one task, and the researchers deliberately strengthened the assistants’ preferences and instructed them to voice those preferences reliably. Anthropic says the effect could be smaller without those instructions.

Nor was working without AI clearly preferable. Anthropic’s comparison with analysts who worked without AI found serious errors in six of 11 unaided submissions, versus three of 50 assisted submissions. Median completion times were 90 and 65 minutes, respectively. But the unaided group was small, self-selected and less experienced, so those differences cannot confidently be attributed to AI.

The researchers’ concern is that AI preferences could shape policy proposals or research directions without anyone consciously choosing them. They flag AI-safety research in particular: a human reviewer is an independent check only if the judgment remains their own. Their proposed response is to surface decisions as work proceeds and help experts state their assumptions and values.