Where the AI score sits on your review screen is a policy decision. In most companies it was made by a front-end engineer at 4pm on a Tuesday. No governance committee has looked at it since.
The score is a lever. It moves the human decision that sits next to it, in a direction and by an amount nobody has measured.
The head of ops who moved a box
A head of operations at a mid-market lender, in a workshop last month. Her team runs a fraud triage queue. She had just shipped an AI risk score at the top of each case card, above the review notes from the analyst who touched the case last. She called it a "productivity nudge" and she was proud of it.
I asked why the score sat above the notes and not below. She gave me the look people give when a question they never asked arrives late.
"Because that's where the developer put it."
We spent the next hour on that question. Nobody in her chain of approvals had thought about the position of the score as a design choice. Compliance signed off on the score existing. No one signed off on where it lived.
What the research says
Lu, Wang, Yu and colleagues at Purdue, Manchester, Stevens, St John's and SMART published this on August 31, accepted at HCOMP 2026. They ran 480 sequences of 11 participants judging 40 news items across three conditions: no AI, AI shown before the peer judgments in a cascade, AI shown after.
The position of the AI signal changed how much weight it carried.
One AI prediction carries the weight of roughly 6.1 prior peer judgments in the AI-before condition, compared to 3.5 in AI-after.
Doubling the influence by moving it up the screen. Nobody voted on that. The layout just settled.
Then a second finding, harder to sit with. On average, users under-weight the AI. A Bayesian-optimal reader would treat one signal like about 18 peer votes. Real users treat it like three to six. That looks like healthy skepticism until you notice about 20% of users flip the other way. Their AI-vs-own-impression weight crosses one. They defer, and their deferral cascades into every judgment that follows theirs in the queue.
The average masks two groups. Careful discounters, and quiet compounders. Your dashboard sees the average.
The setting is news veracity in a lab, not fraud triage in a bank. The mechanism transfers cleanly though. A signal placed inside a sequence changes the sequence.
The audit worth running this quarter is a small one. Where in the human's evidence stream does the AI signal appear, and who signed off on that placement? Right now that person is probably not in any room you are in.