Chinn and Brewer's Seven Responses: A Doubt Sent Forward Has to Be Written── Held inside, a mismatch stops both the learning and the checking; written down and queued for verification, being taught and doubting run together

Material drafted by generative AI lands on your desk, something in it does not fit how you were taught to read, and you leave the room without saying so. When you are being taught something, can you learn it and doubt it at the same time? You can. But a doubt kept inside the head is gone by the next day; it turns into learning only once the mismatch is written down with a person and a date attached to checking it.

01Doubt does its work only once it is written down

In 1993 Clark Chinn and William Brewer sorted the ways people handle data that does not fit their existing beliefs into seven responses. Only one of the seven sends the data forward without a verdict. The other six close the matter where it stands and send nothing on.

A mismatch you leave the room without voicing falls into one of those six. Did you ignore it, reject it, or set it aside as outside your current theory? The person who felt it cannot say which, and the feeling itself goes.

The one placement that sends a mismatch forward works only when the mismatch leaves your head. Paper or a logged record will do equally well, but the inside of a skull will not, because nobody reads it the next day.

The unease you feel about material drafted by generative AI divides at the same point. Write it and the next person can check it. Leave it unwritten and both the drafter and the reader move on without knowing anything was there.

02Five of the seven responses are ways of discounting the data

Here are the seven. Ignore the data. Reject it. Exclude it from the theory you currently hold. Hold it without a verdict. Reinterpret it. Make peripheral changes to the theory. Change the theory itself.

Accepting the data and changing the theory is one of the seven. Holding it without a verdict is the only placement that sends it forward without discounting it. The remaining five all move the data so the belief can stay.

The list is useful because it lets you say afterwards which one you did. "I had a bad feeling at the time" fits all seven equally well. Which one it actually was is settled only by what you wrote down.

One limit: Chinn and Brewer were writing about science instruction and how people learn science. Carrying their framework across to promotional material review is a stretch of my own, not a correspondence anyone has measured.

03In promotional material review, doubt moves only as far as it is written into the record

That which one it was is settled only by what you wrote can be tested where a form for records already exists.

Japan's guideline on sales information provision activities requires that information cited in promotional materials carry an explicit source, and that work records, including records of oral explanations, be created and retained. In review, a doubt has force only to the extent it enters that record. Someone says "this bothers me a little" in a meeting, and if it does not reach the record, by the following week it never happened.

Where generative AI produced the draft, none of this shifts. The one thing that changes is that the party being doubted is not a colleague. When an AI output contains an error, the name left in the record is that of the person who approved the output. What was doubted before approval can be shown only to the extent it was written.

04Sperber and colleagues argued that doubt is a built-in faculty

If only what enters the record moves, it is worth asking where the unwritten doubts came from and where they went.

In 2010 Dan Sperber and colleagues wrote that humans depend massively on communication with others, which leaves them permanently exposed to being misinformed, accidentally or deliberately. For that reason, they argued, cognition carries a suite of mechanisms for filtering what is communicated. They called this epistemic vigilance.

On that account, doubt is not a special posture. It is a default faculty, already running while you are being taught. And a faculty that runs with nowhere to deposit its output ends inside the head that ran it. What is in short supply is not the capacity to doubt but a route by which its output leaves the head.

On that same account, the point of loss can be narrowed. The noticing works. What fails is the step at which there is nowhere to put what was noticed.

05In Buçinca's experiment, putting a step back in reduced overreliance

That things are lost where there is nowhere to put them has also been measured on the tooling side.

In 2021 Zana Buçinca and colleagues tested designs that insert a step before the judgement, such as requiring people to commit to their own answer before seeing the AI system's. Those designs reduced overreliance on AI more than simply showing an explanation alongside the answer. So far, the wanted result.

And yet participants gave the lowest ratings to the designs that cut overreliance the most. Within this experiment, the more effort a design demanded, the lower it was rated. A design that keeps drawing low ratings is less likely to be chosen.

Overreliance itself is not a new finding. A systematic review by Kate Goddard and colleagues in 2012 gathered 74 studies on excessive reliance on decision support. The 74 is the number of studies included, not a rate at which reliance occurs.

The experiment ran on a limited set of tasks. Whether the same effect appears in promotional material review has not been measured.

06Whether doubt survives depends on the form, the sequence, and the appraisal

The finding that the more demanding designs were the least liked splits the conditions for keeping doubt into three.

First, with no field to write in, the doubt does not survive. What changes is whether it reaches the next day at all. This follows from the seven responses: the one placement that sends a mismatch forward works only once it has left the head.

Second, a doubt that survives but never enters a sequence for checking does not move. What changes is whether anyone is fixed to check it. The guideline requires records to be created and retained; it does not say when or by whom they should be read again. Each organisation settles that.

Third, where the person who raises a doubt is rated poorly, nobody raises one. What changes is whether they do it a second time. If the design cutting overreliance most drew the lowest ratings, appraisal of the person who raises a doubt may move the same way. Nobody has measured that with people. In a workplace where whoever halts the work becomes tiresome, neither the form nor the sequence gets used.

07A classroom that bans AI gives up its own training ground for doubting

Of the three, the direction of appraisal is the one that does not change by being specified, as a form or a sequence can be. In places that train people, it shows up first.

Several American law schools have restricted devices in the classroom and the use of AI in submitted work. The University of Chicago Law School announced a ban on laptops, tablets and phones in first-year classrooms, and UC Berkeley Law prohibits students from using AI in conceptualising, outlining, drafting, revising, translating or editing any work submitted for credit.

There is an argument on the other side. An opinion piece in BusinessToday holds that the limitations of AI, confidentiality problems, verification of authorities and recognising doctrinally wrong answers should form the curriculum itself. That is a position, not a measured result.

No study comparing which approach works better turned up in my search. What is plain is that choosing not to let students use the tools removes that much practice at noticing a wrong output and writing it down.

You can be taught and doubt at once. The two hold in the same hour.

Kept inside, though, the doubt falls to the discounting side of the seven, and the fall goes unnoticed too. To keep it, write the mismatch in one line, attach the source of the statement, and put down who checks it and when. Only with all three written does a reader appear the next day.

What is left at the end is how the person who wrote it is treated. Where whoever halts the work becomes tiresome, no amount of tidy forms will produce that one line.