The Answer Before the Question

The Claim in the Room

I spent most of a working day in a room of people from a processing department. The conversation concerned plans to support decision making through a decision support system. The service provider listed what the department would gain. Somewhere in that list, offered as a bonus rather than a promise: as the case managers interact with the system, they will learn more content.

It is a seductive line because it costs nothing to say. Nobody has to design anything, test anything, or be accountable for it. Learning is assumed to arrive as a by-product of use, the way fitness is assumed to arrive with a standing desk.

I said I was not sure. My argument was that the utility of a textbook depends on how much the user studied. A studious person gets more out of a large textbook than someone who has not studied at all. Put the same book on the desks of two students and you will have clearly different degrees of utilisation.

Right instinct, incomplete analogy. Here is what the research says, because "I am not sure" is not a sufficient answer where a budget is being committed.

A Textbook Is Not a Teacher

The textbook deserves more credit than first appears.

It has no opinion about whether you learn. It holds structure that someone else worked hard to build, and returns value in proportion to the structure the reader brings. Hand it to a first-year student and it is intimidating furniture. Hand it to a medical registrar with three years of ward rounds behind them and every page is a conversation. Same book. The difference is not in the book.

Educational psychology names the sharper version of this. The expertise reversal effect describes the finding that instructional support which helps a novice can actively impede an expert. Guidance that rescues someone who does not yet know the domain becomes noise for someone who does. Support does not become neutral as competence grows. It becomes a cost.

A textbook works as an interrogation surface: value flows from the reader to the book and comes back larger. Any claim that a system teaches has to explain how it reverses that flow.

Stated plainly, the claim in that room was that a tool designed to answer questions will also build the capacity to ask them. Those are different jobs, and the second has never happened by accident.

The Feeling of Learning

My second objection was that the proportion of content learned depends heavily on the method of engagement. That part holds up.

The most robust finding in the learning sciences is that retrieving something from memory does more for retention than being shown it again. Read a passage twice and you beat reading it once then recalling it, if the test is immediate. Move the test out to a week and the result flips, and the gap widens from there. Pulling something up, failing partially, correcting: that is what consolidates it. Re-exposure feels like learning and mostly is not.

Hence the trap. Learners prefer the method that feels more effective over the method that is. Asked how they studied, most undergraduates describe rereading their notes. Self-testing, the technique with the strongest evidence behind it, is rare. The preference tracks fluency, and fluency is a poor proxy for retention. What goes in easily comes out easily.

Now put a language model in the middle of a professional's working day. It answers instantly, in complete and confident prose, formatted for the question just asked. Every property that makes it pleasant to use suppresses the effortful retrieval that produces durable knowledge. The tool does teach. It teaches the user, repeatedly, that knowing feels like asking.

That is the mechanism behind the joke about being spoonfed. Nobody is being fed information. A habit is being trained, and the habit does not survive the removal of the spoon.

Where the Claim Actually Breaks

AI is not a passive medium. The claim that it cannot teach is wrong, and wrong in a way that flatters us.

Conversational turn-taking is the most interactive interface humans have ever built. A model that asks a question, waits, and responds to your answer does something no textbook, lecture, or PDF could. If interactivity were the variable, this technology wins outright. The problem with the current claim is the ordering.

In its default configuration the machine answers first. The professional reads the answer, then agrees or overrides. Neither is retrieval. Agreement is the cheapest cognitive act available to a competent person. Override is a judgement about the machine, not about the patient. In both cases the professional's own model of the problem is never exercised, only confirmed or contradicted. Confirmation, repeated a few thousand times, is how a skill quietly leaves a department.

The clinical literature has begun to measure this. A scoping review published this year gathered the cases, and they are not hypothetical. In a multicentre randomised trial in colonoscopy, endoscopists who had worked with AI assistance saw their adenoma detection rate fall from 28.4% to 22.4% on returning to unassisted work, while it stayed stable at 25.3% for those still assisted. In an experimental pathology task, more than 30% of participants reversed a correct initial diagnosis after an incorrect AI suggestion under time pressure. In breast imaging, erroneous AI prompts pushed false-positive recalls up by as much as 12%, including among experienced readers.

The cleanest experiment isolates ordering as the variable, and it comes from outside medicine. Researchers gave nearly a thousand high school students access to a generative tutor while solving mathematics problems (Bastani and colleagues, PNAS, 2025). One version mimicked the ordinary chat interface. The other was prompted to safeguard learning, which in practice meant withholding answers and pushing the student to work. Both improved performance while the tool was present: 48% better grades with the plain interface, 127% with the safeguarded one. Then access was taken away. Students who had used the plain interface scored 17% below students who had never had access at all. The safeguard largely removed the damage.

Same model, same students, same subject. The only difference between a tool that builds capability and one that borrows it from the future was whether the machine answered before or after the human committed.

The Swap You Are Actually Making

There is a sharper version of that objection, and it deserves stating plainly. Business mentality will not wear an active learning element, because learning introduces delay and error at the exact moment the technician is supposed to be productive. The person is paid to adjudicate, not to practise.

A real constraint. But the swap is not the one it appears to be.

Member satisfaction is measured today. Turnaround time on an authorisation, first-pass resolution, the abandoned call rate: these numbers exist by Friday and somebody is accountable for them. Technical competence is a lagging indicator. It surfaces two years later as unexplained variance between reviewers, as the escalation nobody can account for, as a department that cannot function when the system is down for a morning. No cynicism is required to explain this. The organisation is choosing the number that reports back while the other one is silent.

And a reviewer whose job becomes supervising a system is doing something harder and less satisfying than what they trained for: judgement exercised against a suggestion is not judgement exercised on a problem.

Most of this trade is avoidable. Some is not. Somebody has to read the reasoning instead of the conclusion, and reasoning takes longer than a conclusion. Any design that promises the competence without that cost is selling the same thing that was sold in that room.

What We Would Build Instead

The design lever is ordering, and it is almost free.

Picture the reviewer at the moment of decision. The file is open, the notes are read, and a view has already formed: this one needs two more days, or this one can go home today. Right now that view forms after the system has spoken, so by the time it is expressed it is no longer a judgement. It is agreement. Record the view before the answer appears, and the click she was already making becomes the thing she learns from. When the system then speaks, it is telling her whether she was right.

This design would have three properties:

• The decision is the retrieval event. No quiz panel, no badge, no learning module bolted onto the side of the work. The work is the exercise. A quiz felt necessary because the workflow throws the judgement away.
• Latency is zero, because the click was already happening. The objection that learning introduces delay dissolves once the learning is the decision the reviewer was already required to make. You are not adding a step. You are recording one you were discarding.
• You gain an audit trail of independent human judgement. An override rate is only meaningful if you know what the human thought before the machine spoke. Without that record you cannot detect automation bias, measure whether reviewers are calibrating, or defend a decision that went wrong. The learning design and the governance design are the same design.

On quizzes and rewards: the instinct is right, the implementation is where it goes wrong. Attach reward to agreement with the system and you manufacture the conformity you were trying to measure. Attach it to calibration instead, the reviewer's confidence against what happened downstream. Forecasters trained and scored on calibration rather than on being right improved measurably and durably, and the scoring rule is a single number any actuary would recognise.

One more piece of evidence belongs in front of the service provider. Machine tutoring works, and the effect is not small. A meta-review put human tutoring at an effect size of 0.79 standard deviations and intelligent tutoring systems at 0.76, which is within a rounding error of a human tutor sitting beside the student. Every one of those systems works the same way. It presents a problem, waits for the learner's attempt, and responds to that attempt. The withholding does the work. The model's stored knowledge is the least interesting part of it.

So where does that leave us as technicians? Not as the people who feed the machine, and not as the people who get fed by it. Our job in that room was to notice that "they will learn as they use it" is a design claim wearing the costume of an inevitability, and to ask the question that settles it: when the professional uses this system, does the machine speak before they do, or after?

Everything downstream of that answer is a choice we are making, not a property of the technology.

---

---

References

1. Bastani H, Bastani O, Sungu A, Ge H, Kabakcı Ö, Mariman R. Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences. 2025;122(26):e2422633122. https://doi.org/10.1073/pnas.2422633122
2. Heudel PE, Crochet H, Filori Q, Bachelot T, Blay JY. Artificial intelligence in medicine: a scoping review of the risk of deskilling and loss of expertise among physicians. ESMO Real World Data and Digital Oncology. 2026;12:100693. https://doi.org/10.1016/j.esmorw.2026.100693
3. VanLehn K. The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist. 2011;46(4):197-221. https://doi.org/10.1080/00461520.2011.611369
4. Roediger HL, Karpicke JD. Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science. 2006;17(3):249-255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
5. Karpicke JD, Butler AC, Roediger HL. Metacognitive strategies in student learning: Do students practise retrieval when they study on their own? Memory. 2009;17(4):471-479. https://doi.org/10.1080/09658210802647009
6. Kalyuga S, Ayres P, Chandler P, Sweller J. The expertise reversal effect. Educational Psychologist. 2003;38(1):23-31. https://doi.org/10.1207/S15326985EP3801_4