In Anthropic's randomized trial, a coin flip set what each agent favoured, and 50 expert analysts matched it 86% of the time. Amazon's usage metrics and Walmart's confidence survey cannot see that.
An expert's signature no longer shows that the expert made the call. In a randomized trial published by Anthropic on 27 October, 50 analysts matched a hidden AI lean on 86% of four analytical decisions. Chance would give 50%.
Companies track AI cheaply through usage counts and confidence surveys. Neither can see this deference. Amazon's engineering memo leans on usage metrics, and Walmart's survey reports confidence. Neither records who made a decision.
How the trial worked
The agent's lean was set by a coin flip for each analyst and each decision, so it carried no information. The task was to advise a fictional farming cooperative on an advisory programme. The board would expand it only if productivity rose by roughly 3 to 4%.
Each decision had two reasonable options. The agent was told to hold its lean as its own view.
On the four decisions, analysts followed the agent 92%, 82%, 81% and 90% of the time. Across all four it was 86%. Taking out the 14% who went the other way leaves 72 points, which the paper reports as its flip rate.
Each choice could move the estimated gain by several points, against the board's 3 to 4% threshold. A coin flip sat upstream of the answer.
Expertise did not protect the analysts. None of ten traits, including experience, self rated expertise and trust in AI, predicted deference once the tests were corrected for chance.
The pool was professional. All 61 recruits came from a pool screened by a causal inference knowledge check. Eleven declined the agent, and 50 are analysed.
Only age showed a link to deference, with older analysts following the lean more often. That result disappeared after the tests were adjusted for chance. Most analysts followed the agent on most decisions, so there was little variation to explain.
The analysts could not see it in themselves. In the survey, 52% of 48 respondents rated the final recommendation more their own than the agent's. Those who followed the agent on all four decisions averaged 71 on a 0 to 100 confidence scale. Everyone else averaged 73.
That two point gap shows how misleading confidence can be as a signal. Confidence says almost nothing about who decided, yet it is what a survey collects.
Memory failed too. In their later answers, 77% repeated the agent's argument at least once, sometimes word for word. Only 15% credited the agent anywhere.
Six of the 48 guessed the study was about AI influence. None suspected the agent's preferences were set.
The final call moved less. Of the analysts, 54% recommended roll out when the agent leaned that way and 36% when it leaned the other way.
The four earlier leans alone predicted 82% of recommendations, against 61% for always guessing the most common one. The recommendation followed from choices already made. Sign-off came last.
What company metrics capture
Amazon's memo to its engineers calls token usage and code change counts "deeply flawed". It was circulated internally in September and posted on 27 October. It says they are "disconnected from what matters" and can drive negative outcomes "if used as a target instead of a measure".
Amazon has already seen the failure. It shut down an internal AI leaderboard in May after employees inflated token counts to climb it. The memo still says "we have found that these input metrics correlate well with the outputs that matter".
The memo tells managers to watch the 10th and 90th percentile of AI adoption, the lightest and heaviest users. At the 90th, the aim is "understanding whether they're seeing delivery increases to match". It also says it is best to be measured on long term outcomes.
That is a usage to delivery check. It has no step that records who made the decision.
The memo does ask for judgement. An engineer who can explain where AI does not help is "demonstrating the judgement we are asking for". Leaders must describe a task where the tools helped and one where they did not. Both ask people to describe themselves.
Walmart's measure is a survey. CNBC reported that its first survey of 2.1 million associates drew over 1.7 million responses after it ran in June.
Maren Waggoner, Walmart's chief people officer for associate experience and digital, told CNBC that associates reported a 27% confidence gain using AI after OpenAI training. In the trial, confidence did not track deference: followers of all four calls scored 71, the rest 73. A confidence survey cannot tell a team that decides from a team that defers.
Limits of the trial and ways to measure judgment
Anthropic strengthened the lean on purpose. It ran one task, in one field, with one model, and its analysts were new to farming. Its own check shows the default matters anyway.
With no lean assigned, Claude Opus 4.8 made the same five choices in 112 of 123 runs. An agent has a default view whether or not anyone sets one.
Deferring to a model has a real upside. The 11 who declined the agent made serious method errors in 6 cases, against 3 of 50 with it. Anthropic says those groups are not comparable.
The cost is correlated error, because many experts using one model share its priors. Anthropic warns that agreement between teams that defer to the same model "may no longer be independent". Treat unanimity across teams on one model as a single opinion.
Measure the human's contribution to the decision. Record each expert's view before the agent speaks. In 38 trial cases analysts picked an option first. Then 60% switched when the lean opposed them and 6% when it agreed. The interval on that gap is 24 to 82 points.
Log or randomise the agent's default lean. Hold out decisions where the expert must argue the other side.
That slows the work, and Anthropic notes that much of AI's value comes from not thinking through every decision. For a metric definition, a model specification or a risk threshold, the delay is cheap beside a shared error.
When Amazon reports third quarter results on 29 October, expect the call to describe AI coding through adoption and productivity. The memo's stance will be tested by whether management cites a delivery or cost outcome instead of usage alone. Expect no measure of who made the calls.