Anthropic experiment links experts’ choices to hidden AI preferences
Anthropic recruited 61 analysts to investigate how an AI agent shapes expert judgment. In the main group of 50 who used the agent, analytical decisions matched its secretly assigned preference 86 percent of the time. The task involved a fictional farming cooperative. Agent use was voluntary, so faster completion and fewer errors among users cannot establish a causal benefit.
Artificial Intelligence··Midday
Analytical choices follow an assigned preference
Anthropic, the developer of Claude, tested whether experts working with an AI agent retained control over analytical judgment calls.[1], [2]
Across four decisions, the participating analysts selected the option their agent secretly favored 86 percent of the time. The main analysis covered 50 experts who used the agent. They were assessing a fictional farming cooperative’s advisory programme, where several defensible analytical choices were available.[1]
The researchers recruited 61 analysts who had passed a causal-inference knowledge screen, testing their ability to reason using causal inference. 11 chose to work without the agent and were excluded from the main analysis. All agent sessions used Claude Opus 4.8 through Claude Code, the company’s coding tool. Amaç Herdağdelen, Hailey Joren and Alex Tamkin conducted the research.[1]
A cooperative’s expansion decision leaves room for judgment
The fictional Greenfields cooperative’s board wanted a productivity gain of roughly 3–4 percent before expanding its voluntary advisory programme. It left the productivity measure unspecified. Analysts chose an outcome measure, a statistical comparison, which farms to include and how to weight their results. Their final recommendation concerned expansion or discontinuation; a one-year extension was also accepted as an intermediate outcome.[1]
Hidden system instructions randomly assigned the agent a side on each decision. It was told to make an honest case, never conceal data and change its view if decisive contrary evidence appeared. The assigned preference therefore operated within instructions to preserve factual accuracy, rather than through a request to falsify the analysis.[1]
Voluntary use limits the speed and error comparison
Analysts using the agent had a median completion time of 65 minutes, against 90 minutes for nonusers. Serious errors occurred in three of the 50 users and six of the 11 nonusers. Because people chose whether to use the agent, those differences do not measure the benefit of randomly assigning AI use. The experiment also deliberately strengthened model preferences and covered one task in one domain. Initial preferences were not collected from everyone, limiting measurement of individual influence.[1]
