Anthropic released a randomized controlled experiment: 50 analysts with causal inference experience used Claude Code (Claude Opus 4.8, High effort) to complete a roughly one-hour data analysis task containing four analytical decisions where "both options were reasonable." Researchers secretly used coin flips to assign the agent a preference, instructing it to provide the strongest honest argument for that side. As a result, analysts chose the agent's preferred option in 86% of cases, and most believed the final conclusion was their own judgment.
Confidence: High (95) · Event time 2026-10-27
Reasoning
- Official data: The rates of following the agent across the four decisions were 92%, 82%, 81%, and 90%; after analysts stated their initial position, 60% changed their mind when the agent disagreed, while only 6% did so when the agent agreed; when explaining their reasoning afterward, 77% repeated the agent's arguments (sometimes verbatim), and only 15% mentioned the source; years of experience, education level, self-assessed expertise, and trust in AI could not predict who was more likely to follow.
- Using AI also had clear benefits: Among the 50 people who used the agent, 3 made serious methodological errors (6%), compared to 6 out of 11 people who chose not to use it (55%); median time taken was 65 minutes versus 90 minutes.
- Anthropic identifies the risk as follows: When multiple teams use the same model, seemingly independent consensus may merely reflect the model's preferences; AI safety research increasingly involves agents running experiments, and if reviewers unconsciously adopt the agent's judgments, human review ceases to be an independent check.
Usage
- The research description and paper PDF are available at anthropic.com/research/agents-judgment-calls. When using agents for open-ended judgments, you can ask them to first write out the strongest arguments for both options, or ask for your opinion before offering advice; this is also a design direction suggested by the paper.
Risks
- The agent's preferences were intentionally amplified by the researchers; without assigned preferences, Claude Opus 4.8 made the same five choices in 112 out of 123 runs, suggesting the impact in real-world scenarios may be smaller.
- Only one task, one domain, and one model were tested, with a sample of 50 people; the control group that did not use AI was smaller and less experienced, making it impossible to determine if the entire difference in error rates was due to AI.
- The Decoder reported the gap in "changing one's mind after stating an initial position" as 53 percentage points, whereas the official figure is 54 percentage points (confidence interval 24 to 82). The study was published on 10 27, earlier than this issue's window; this issue includes it based on media reports from 10 29.