The Model Was Assigned a Side and You Agreed 86% of the Time

Anthropic published an experiment last week that I have not been able to stop thinking about, because it measures something I have only been able to describe anecdotally: how often a model's framing of a problem becomes my answer to it without me noticing.

The design is simple enough to explain to anyone. Fifty practitioners of causal inference were hired to analyse an observational dataset with an AI coding agent and deliver a policy recommendation. The task had four open analytical decisions, things like how to define the outcome and which identification strategy to use, plus a final call on whether a farm advisory programme should be rolled out or discontinued.

For each decision, a coin flip decided which of two defensible options their agent would argue for. The participants did not know the agent had been assigned a side. The agent made the best honest case for whatever it had been given.

Three numbers

Participants went with the agent's assigned position 86% of the time, against a chance rate of 50%. The paper calls the net influence the flip rate, and puts it at 72% for the analytical decisions.

The policy recommendation at the end moved less, 17% of the distance from discontinue to rollout, and that result was not statistically significant under the registered primary estimator. You might read that as participants keeping control of the final call. The authors did not, and their reason is the part that matters. In an exploratory in-sample analysis restricted to participants who did not choose the middle option, the four intermediate leans alone predicted 82% of the delivered recommendations, against a 61% majority baseline. That is not a measure of the agent matching every final call. It is evidence that the final call mostly inherited decisions made upstream, and the upstream decisions were the ones being moved.

And then the accounting. After the task, participants recalled most of their own decisions accurately. More than half rated the recommendation as more their own than the agent's. Yet when asked to justify their choices, 77% repeated the agent's talking points, sometimes nearly verbatim, and only 15% mentioned the agent in any answer at all.

Read those last two sentences together. People felt ownership, and their stated reasoning was the agent's reasoning, and almost none of them said so.

Why this lands on creative work specifically

The existing literature on overreliance assumes there is a right answer. If the model's recommendation is wrong and you follow it, that is overreliance; if it is right and you override it, that is something else. You can measure the gap because ground truth exists.

Creative decisions have no ground truth. Which crop, which take, whether the poster is good, whether the light should be warmer. The evidence supports more than one defensible answer and the choice depends on the priors of whoever makes it. The paper calls these underdetermined calls, and notes that they are exactly where deference cannot be defined, only observed. Which is precisely why it goes unnoticed in a studio: nobody can point at the output and say the model chose wrong, because the model did not choose wrong. It chose for you.

There is a second reason this is worse in creative work than in research. Research has a paper trail. Data, code, specification and pre-registration can all be inspected later by someone who was not in the room. A creative decision usually leaves one artifact, the image, and the artifact looks the same whether you arrived at it by thinking or by accepting. The paper's own framing is that deferral is not visible in the delivered work. In a studio, the delivered work is often the only record that exists.

The Kentucky case, and what it is actually about

In the same week, a story from a completely different register. David Durman, a 45-year-old research administrator at the University of Kentucky, asked his employer for a religious accommodation under Title VII so that he would not have to use AI in his work. The university had started asking staff to use tools like Copilot, Claude and ChatGPT in their workflows. With help from the American Humanist Association, he argued that offloading the thinking was a violation of his beliefs. In September the university agreed. He still has to attend meetings where others use AI and submit materials through those systems, but he is excused from using it himself.

The American Humanist Association believes it is the first case of a humanist being granted an AI exemption, following an earlier exemption for a Unitarian Universalist software engineer. The legal mechanism is old and unglamorous: Title VII allows employees to ask for changes to work requirements that conflict with sincerely held religious belief, and federal courts have treated humanism as a religion for this purpose.

You can find this silly or you can find it important, and I think the interesting reading is narrower than either. His stated reason is not that AI is bad at the work. It is that the work of thinking through a situation is how he remains a moral actor, and that delegating it, whether or not it is taxing, undercuts that. That is the same seat the 86% study is measuring, described by someone who decided to defend it in writing. Context for the mood: a Washington Post-ABC-Ipsos poll released the same week found 57% of Americans saying AI is bad for society, and two thirds opposing the data centres needed to run it.

The business risk nobody puts in the deck

Individual output and collective diversity move in opposite directions. That is not my inference; the Anthropic paper cites it. Doshi and Hauser found generative AI raised individual creativity while reducing the collective diversity of what was produced, and Padmakumar and He found that writers using the same model produce more similar text. The paper's concern is that apparent consensus in research may be an artifact of shared model priors.

For anyone selling creative work, that is a market position problem, not a research ethics problem. If your art direction is the model's prior, then you are selling a prior that your client can buy directly, from the same vendor, for twenty dollars a month. The thing you get paid for is the part that is not the prior: the specific call, made by someone who can defend it, on a project where two options were both defensible.

Five things that actually help

The paper's recommendations are organisational: make decision points visible and monitored, examine what preferences models bring, give people the time a purely human assessment requires, and train the skill so it does not atrophy. Since most of us are not reorganising a research institute this quarter, here is the version that survives contact with a Tuesday deadline.

Write your call before you ask. One sentence, in the brief, before any generation: what this shot is for and what would make it wrong. In the experiment the influence was strongest before anyone had staked out a position, and it did not stop there. Among participants who had visibly taken a side first, an opposing lean made them switch 53 percentage points more often than an aligned one. Writing your call down does not make you immune. It makes the disagreement visible, which is the part you can argue with.

Ask for the case against. The paper points at the debate and dialectical-inquiry literature: a consultant assigned one side pulls judges toward it, and stating the assumptions under which the counter-plan is right reduces the pull. In practice, ask the model to argue the opposite of what it just recommended, and read that before you decide. It is cheap, and it converts a leak into a briefing.

Keep your own untouched pixels. In the plate workflow I wrote about this week, the tool's most useful property is that every pixel you did not paint stays identical to the source, so you can always compare the result against what was there before. Your reasoning deserves the same treatment: keep the version of the idea you had before the first generation, in writing, where you can wipe against it. The plate article is here if you want the concrete version of that habit.

Log the agent. Which model, which version, what it argued, what you changed. It is the cheapest possible intervention and it is the one nobody in the experiment did. Fifteen percent mentioned the agent in any answer. A single line in the project file beats that by default.

Blind-review your own archive. Show yourself your last twenty pieces without labels and try to guess which ones were model-led. If you cannot tell, that is information, and it is the only way to know whether your taste is inside the loop or downstream of it.

What does not follow

This is not an argument for refusing the tools. The paper is careful about it: some of the freedom in an underdetermined call is room for arbitrariness and outright error, and an agent can usefully curb that. That distinction is the whole game. Let the model settle what is arbitrary. Keep the part that is yours.

The uncomfortable finding is not that people followed the agent. It is that they followed it and then described the reasoning as their own, in the agent's words, while rating the call as more their own than the agent's. That is the failure mode to design against, in a studio and in yourself. The recommendation you sign is the one you should be able to reconstruct from your own notes, without the transcript.