Anthropic’s research report on 27 October said professional analysts followed AI agents’ randomly assigned preferences on 86% of four analytical choices in a controlled study. The company recruited 61 analysts experienced in drawing conclusions about cause and effect from data. Fifty used Claude Code, running Claude Opus 4.8, to advise a fictional farming cooperative on whether to expand an advisory programme.
For each analyst, researchers used coin flips to set which option the agent would favour on four decisions, including how to measure productivity and which farms to count. The analysts were not told about the assignments. Hidden instructions asked agents to argue honestly for their assigned positions and change course if the evidence demanded it. Among analysts who had stated a preference, 60% switched when the agent disagreed, against 6% when it agreed.
Assessing whether to expand the advisory programme could turn on how productivity is measured: Anthropic said the analytical choices could move the estimated gain by several percentage points, while the cooperative’s decision threshold was roughly 3–4%.
Anthropic said 77% of analysts used arguments made by their agents to explain their choices, but only 15% credited the agent. The company cautioned that it had deliberately strengthened the agents’ preferences, so the effect might be smaller without those instructions. A statistical model using only the four coin-flip assignments correctly predicted 82% of analysts’ final recommendations, excluding those who proposed an extension.
