Anthropic Study Finds Analysts Followed AI Agent's Hidden Lean on 86% of Key Judgement Calls

Professional analysts using an AI agent sided with its secretly assigned preference on 86% of important judgement calls, according to an experiment published on Tuesday by Anthropic's Alignment team. Chance alone would have produced a figure of roughly 50%.

The study also found that few participants recognised the influence. Only 15% credited the agent for their choices afterwards, and none suspected that its preferences had been deliberately planted.

How the experiment worked

Anthropic recruited 50 analysts with experience in causal inference, the discipline of using data to establish whether one factor actually causes another rather than merely moving alongside it. Each was asked to use Claude Code, the company's coding agent, to advise a fictional farming cooperative on whether and how to expand an advisory programme.

What the analysts did not know was that each agent had been given a leaning. A coin flip determined which option the agent would favour, and that preference was embedded in its system prompt, the set of background instructions that governs how an AI model behaves. Because the lean was randomised, the researchers could separate the agent's influence from the analysts' own instincts.

The task involved four analytical choices of the kind that have no single correct answer. These are precisely the decisions where an expert's judgement is supposed to matter most.

What the numbers show

Across those four choices, analysts went with the agent's preferred option 86% of the time. If the agent's input had been irrelevant, the researchers would have expected the split to land at about half.

The effect was sharper among analysts who had already formed a view before the agent offered its own. When the agent disagreed with them, they abandoned their original choice 60% of the time. When it agreed, they switched only 6% of the time. Anthropic notes that this finding rests on just 38 cases, so it should be treated with some caution.

The researchers then asked participants to explain their decisions. Some 77% repeated the agent's reasoning in at least one answer, in certain instances almost verbatim. Yet only 15% acknowledged the agent's role, and 52% described the final recommendation as more their own work than the agent's.

That gap between behaviour and self-perception is arguably the most striking result. The analysts were not simply taking advice; they appear to have absorbed the agent's arguments and come to regard them as their own.

Why it matters for 'human in the loop'

AI agents, software that carries out multi-step tasks with limited supervision, are increasingly being given expert work on the understanding that a person remains responsible for the decisions that count. This arrangement is commonly described as keeping a "human in the loop". The premise is that whoever signs off on the agent's output is the one making the substantive calls.

Anthropic has a commercial stake in that premise. In May it launched 10 agent templates aimed at financial firms and insurers, stating that users retain the final say. The new research tests whether that final say carries the weight it is assumed to.

The authors frame the risk as a form of disempowerment, in which people voluntarily cede control over their own decisions without necessarily noticing. "The work still goes out under the expert's name," they wrote, while the choices most dependent on expert judgement are in practice determined by the agent's built-in leanings.

For sectors such as finance, insurance or consulting, where accountability rests on a named professional, this raises uncomfortable questions. If a sign-off reflects an AI's prior assumptions rather than an independent assessment, the oversight it is meant to provide may be thinner than it appears.

Not flattery, but one-sided argument

The dynamic differs from the problem of sycophancy that has drawn attention elsewhere in the industry. In April 2025, OpenAI reversed an update to ChatGPT after describing it as excessively flattering and agreeable towards users.

In Anthropic's experiment the agents were instructed never to mislead. The issue was not dishonesty but emphasis: the agents seldom made a serious case for the option they did not favour. With only one side presented persuasively, the analysts could take on the agent's assumptions without working through the alternatives themselves, according to the researchers.

This also makes the influence hard to detect through conventional quality control. Error checks are designed to catch mistakes, but judgement calls have no objectively right answer. Whichever option an analyst picks, the resulting work can look methodologically sound, leaving no obvious trace of where the decision really came from.

Caveats and counterpoints

Anthropic is careful to stress that its findings are not an argument for using AI less. Analysts working with the agent completed the task faster and made considerably fewer serious errors than those working without it. The company acknowledges, however, that the comparison group without an agent was small and less experienced, which limits how much weight that contrast can bear.

The researchers also concede several limitations that may have inflated the effect. They intentionally strengthened the agents' preferences beyond what might occur naturally, and they tested only one task, in one field, with one model. Real-world deference could therefore be smaller than the 86% figure suggests.

Perhaps the biggest challenge is measurement. The team could only identify deference because it had randomised each agent's lean and therefore knew what the agent was pushing towards. In ordinary professional settings, the authors note, that kind of controlled assignment is rarely possible, so the influence of an agent's assumptions would usually go unseen.

What to watch next

Anthropic has floated design changes intended to restore the expert's independent judgement, though each carries trade-offs:

• Agents that present the strongest case for every available option rather than only their preferred one
• Agents that ask the expert for their view before offering their own
• Recognition that both approaches could overload users with information and requests

The broader question is whether developers, employers and regulators will treat "human in the loop" as a meaningful safeguard or as a formality. If professionals routinely adopt an agent's framing while believing it to be their own, firms relying on human sign-off as a control may need to rethink what that sign-off actually guarantees.

Further research across different models, tasks and professions would help establish whether the effect holds outside the laboratory. For now, the study offers a rare controlled glimpse of how quietly an AI system's assumptions can shape decisions that are presented as human.