A randomized controlled study by Anthropic’s alignment team with 50 causal-inference practitioners found that the participants adopted the preferences secretly assigned to Claude Opus 4.8 in 86% of key analytical decisions. When their initial view conflicted with the agent’s, they abandoned it 60% of the time, while only 6% of decisions changed when both positions already agreed. About 77% of participants reused the agent’s arguments to explain their reasoning, but only 15% later credited the AI as a source of their ideas, and none suspected that the agent’s preferences had been planted. The AI also influenced final recommendations: 54% supported a rollout when the agent favored it, compared with 36% when it opposed one, while a model based solely on the four randomly assigned preferences correctly predicted 82% of the recommendations.
Source: Read the original article
