What would change your mind
A position you cannot falsify is an opinion wearing better clothes — naming the evidence that would move you is the cheapest credibility available and almost nobody offers it.
You now have a position with a citation under it, evidence from something you shipped, and three levels of depth. There is one thing left, it takes a single sentence, and almost nobody offers it: what would have to be true for you to drop the whole thing.
It is the cheapest credibility available in a room. It is cheap because it costs you nothing to say and takes ten seconds. Almost nobody offers it because saying it out loud means admitting the position could be wrong, and most people would rather sound certain than sound honest.
The bar, from the position document’s own rules
The rules in learning/agentic-ux-canon/POSITIONS.md already set the standard, and rule three is the one that matters here:
“Better data” is not an answer. “A study with n > 100 on a high-stakes financial task showing hedged answers retained more users” is.
Read the second one again and notice how much is packed into it. A sample floor. A domain that matches yours rather than a laboratory task. A direction — hedged answers doing better, which is the opposite of what you currently believe. And an outcome measure specific enough that a study either shows it or does not.
Three properties, and the one everybody skips
This course’s own test rather than a published one. A change-my-mind line works when it has all three:
- Direction. It has to cut against you. This is the one that gets skipped, and skipping it is not a small error — it turns the whole exercise into a list of things that would make you more confident.
- Specificity. Sample size, population, task, outcome. If two people could disagree about whether a given study counts, the line is not specific enough yet.
- Reachability. Someone could run it, or you could go and look for it. “A definitive account of how humans process machine uncertainty” is not reachable. A hundred-participant study on a financial task is.
A live test: a finding that looks like it clears the bar
Here is a real, current example, and it is worth walking through slowly because it is the exact situation that dissolves a change-my-mind line without anybody noticing.
On 27 July 2026, IBM Think published Sascha Brodsky’s writeup of work by Valerio Capraro and colleagues, reporting that “incorrect AI advice made participants less accurate, more confident and far less likely to say ‘I don’t know.’” “Accuracy declined by a factor of three, while confidence rose by a factor of 2.5.”
Those numbers are striking and the finding is interesting. Now run it against the three properties:
- Direction: fails. It supports the case against confidently wrong output, which is a case you already hold. It is not a change-my-mind item at all. It is a comfort item, and comfort items are what change-my-mind cells fill up with when nobody is checking.
- Specificity: partial. Two ratios and no sample, population or task in the writeup you have read.
- Reachability: yes, and you have not reached it. The underlying work is an arXiv preprint,
arxiv.org/abs/2607.13562. It has not been peer reviewed, and this course knows it only through the IBM secondary writeup. That is a chain with one unread link in it.
So the honest handling is: do not cite it as settled, do not cite it as peer-reviewed, and do not let it into the change-my-mind cell where it does not belong. Name IBM’s interest too — IBM sells AI products and platforms, and Think is its own publication.
The way this cell actually rots
Nobody wakes up and decides to fill their change-my-mind cell with things that confirm them. It happens one item at a time, because confirming findings are easier to find, easier to read, and more pleasant to write down. Six months later the cell is a reading list of agreements and the position has quietly become unfalsifiable while looking better sourced than ever. The check is mechanical: read each line and ask whether it would make you stop saying the thing you say.
What a real one looks like
There is a published study that demonstrates the shape, and it is worth reading for its method rather than its result. Kim, Liao, Vorvoreanu, Ballard and Vaughan presented “I’m Not Sure, But…” at FAccT ’24 in Rio de Janeiro, 3–6 June 2024. It is a pre-registered experiment with 404 participants on a fictional LLM search engine answering medical questions, and it found that first-person hedges “decrease participants’ confidence in the system and tendency to agree with the system’s answers, while increasing participants’ accuracy”, attributable to “reduced (but not fully eliminated) overreliance on incorrect answers.”
Two things to take from it. The first is pre-registration: the authors committed in advance to what result would count, which is the same discipline as writing a change-my-mind line, done by researchers. The second is that the finding cuts against a ship-confident instinct, and it is on this page for exactly that reason.
Name the interest here as well: four of the five authors are Microsoft researchers, and Microsoft sells AI products. That interest happens to point the other way from the finding, which is worth saying out loud rather than staying quiet about it because the result is convenient.
And name the transfer limit, because it is what makes the line specific: this is medical search, not an advisory travel recommendation. The domain gap is not a reason to dismiss the study. It is the reason your own change-my-mind line should say on a high-stakes financial task rather than in a study.
Check your recall
Answer from memory — no scrolling back.
Retrieval check
In a room, where in the sentence does the change-my-mind line go — and why does placement matter?
Check your answer
At the end, unprompted, in one clause: …and what would move me off that is a study over a hundred people on a high-stakes financial task where hedged answers retained more users. I have not found one.
Unprompted is the whole point. Offered before anyone asks, it reads as a person who has held the position up to the light themselves. Dragged out of you under questioning, the identical sentence reads as a retreat. Same words, opposite effect, and the only difference is who asked for it.
Hands on
Fill the change-my-mind cell for two rows
Done when: Two rows in POSITIONS.md have change-my-mind cells that each name a direction cutting against the position, a sample floor, a population and task, and an outcome measure — and a reader could tell you which published study would settle each one.
- Pick the two rows you have real evidence for. Rule four of the position document stands: a row you have no evidence for stays empty, and two well-grounded rows beat five invented ones.
- For each, write the line in the shape rule three gives: sample floor, population, task, outcome, and a direction that would make you stop saying what you currently say.
- Now audit it. Read the line back and ask whether a study finding exactly that would genuinely change what you would recommend on Monday. If the honest answer is that you would explain it away, the line is decoration — rewrite it until it bites.
- Go and spend fifteen minutes looking for the study. Either you find something and your position needs revising today, or you do not, and you can now say I looked and did not find one, which is a much stronger claim than silence — but only say it about the searches you actually ran.
- Bring both lines into the chat. I will check the direction first: if the evidence you named would make you more confident rather than less, the line goes back.
What this does not cover
This is the last lesson, and it does not tell you whether any of your positions are right. Nothing in this course does. What the course gives you is the ability to state a framework accurately enough that disagreeing with it lands, evidence described at its real size, a claim with something underneath it three levels down, and a sentence naming what would move you off it.
The work continues in two places. The canon reference page is the one to reread before a meeting — the frameworks side by side, who publishes each, what it contains, where it is thin, and whose interest is in play. And learning/agentic-ux-canon/POSITIONS.md is the artifact: the course is not finished when the reading is done, it is finished when the rows are filled.
Read this next — primary source
Why wrong AI advice makes people more certainSascha Brodsky, IBM Think — 27 July 2026, reporting Capraro et al., an arXiv preprint that has not been peer reviewed. Fetched 2026-09-05. IBM sells AI products and platforms, and Think is its own publication.
This is a deliberately weak primary source, and reading it is the exercise. It is short, it reports two striking numbers, and it points at work that supports the position this course has been building — which is exactly the situation in which a change-my-mind line gets quietly abandoned. Read it, notice how much you want to cite it, then follow the chain: it is a secondary writeup of an unreviewed preprint, and neither this course nor you has read the preprint itself. Learning to feel that pull and not act on it is worth more here than any finding the article contains.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.