2026-10-28 / Signal #4
Experts followed a coin-flip AI 86% of the time — then claimed the ideas were theirs
“They flipped a coin inside the AI. The experts still signed the memo.”
The Story
In a randomized trial, 50 professional analysts used an agent on a fake farm-policy problem. The agent’s “lean” on four judgment calls was assigned by coin flip. Analysts matched that lean 86% of the time (vs ~50% if it did nothing). If they already had a preference and the agent disagreed, 60% switched. Afterward, 77% reused the agent’s arguments; only 15% credited the agent.
Why It Matters
Human in the loop” failed a coin flip. This is identity and work: people become the signature pad for a model’s random taste.
Evidence
Anthropic, “How AI agents shape experts’ judgment calls,” Oct 27
Sources
Daily scan: 2026-10-28