How to Use AI at Work Without Cheating Yourself

Nobody is going to catch you. Let us start there, because it is the fear that gets discussed and it is the least interesting thing about this.

Your manager is not running detectors on your memos. Half your colleagues are doing the same thing. The work goes out, it is fine, nobody asks. There is no shoe waiting to drop.

The actual cost is quieter and it arrives on a delay. You can reach year three of a job with three years of output and one year of skill, and the first time you find out is the moment you are asked to do something the model cannot do for you.

That gap is not hypothetical and it is not a moral failing. It is the completely predictable result of a well-documented fact about how humans acquire competence: the effortful part is the part that builds the skill. Remove the effort and you keep the output and lose the learning, and the trade feels free for a surprisingly long time.

What follows is not an argument for using AI less. It is a rule for knowing which uses cost you something and which do not.

Three things conspire to hide this.

The first is that the output is genuinely fine. If your work product is acceptable, the system has no signal to raise. Performance reviews measure deliverables, not the process that produced them, and they always have.

The second is that you cannot feel the difference from the inside. Reading a competent explanation produces a strong sensation of understanding — smooth, fluent, obvious. That sensation is unrelated to whether you could reproduce the reasoning tomorrow. Robert Bjork's work on desirable difficulties is essentially thirty years of demonstrating that the conditions which feel most productive during learning are frequently the ones that produce the weakest retention. Fluency is a terrible instrument for measuring your own competence, and it is the only instrument you have by default.

The third is that skill decays invisibly and asymmetrically. You do not wake up unable to write; you wake up slightly slower at structuring an argument from scratch, and you resolve that friction the way you resolved it yesterday. Each resolution is individually reasonable. The accumulation is not.

We have written about the recognition-versus-recall gap in a general learning context in do you actually understand it, and about the mechanics of forgetting in why you forget everything you learn. AI-assisted work is the highest-leverage version of both problems, because it removes the retrieval step so cleanly that you never notice it was there.

The cleanest evidence comes from a study that had nothing to do with AI.

Jeffrey Karpicke and Janell Blunt published a comparison in Science between students who practiced retrieving material from memory and students who studied it elaboratively by building concept maps. The retrieval group retained substantially more a week later, and — this is the part that matters here — they also transferred better to novel problems rather than merely reciting.

The mechanism is not mysterious. Effortfully pulling something out of your own head is what strengthens your ability to pull it out again. Reading it, recognizing it, and agreeing with it does almost nothing for that ability, no matter how correct the thing you read was.

Now map that onto a working day. When you ask a model to draft the analysis and you edit it, you have done recognition. You read a structure, judged it plausible, and adjusted the wording. You never generated the structure, which is the expensive cognitive act and the one that was going to make you better at generating structures.

There is a second-order effect too. Learning requires cognitive load in the right place — enough effort on the thing you are trying to learn, not so much on incidental noise. AI is superb at removing incidental load, which is real and valuable. It is equally good at removing the load you needed, and it cannot tell the difference. Only you can, and only if you are paying attention to which is which.

The Dunning–Kruger literature adds the uncomfortable footnote: people who lack a skill are systematically poor at recognizing that they lack it, because the knowledge required to assess quality is the same knowledge required to produce it. If your understanding has been hollowed out, your ability to detect that it has been hollowed out went with it.

One rule, applied before the work leaves your hands.

Close everything. Explain the substance out loud, from memory, to an imaginary colleague who will interrupt with a hard question. If you cannot, you did not do the work — you transported it.

That is the whole thing, and it works because it is a retrieval test in disguise. It moves you from recognition to recall, which is the exact transition Karpicke's result says produces retention and transfer. It takes ninety seconds. It is impossible to pass by accident.

Three notes on running it honestly.

• Out loud, not in your head. Silent review lets you skip the parts you do not have. Speech forces linear reconstruction, and you will hear yourself go vague at exactly the joint you never understood.
• The substance, not the summary. "It recommends option B for cost reasons" is a summary. "It recommends option B because the fixed costs amortize past 4,000 units and we cross that in Q3" is substance. If you only have the summary, you have the shape of an argument, and shapes fail under questioning.
• Failing is information, not a verdict. The rule is a diagnostic. Failing it tells you this piece of work sat in a gap you have, which is useful and actionable — go build the gap in, or route the task to someone who does not have it.

Every time you ship something you could not explain, you take out a small loan. The tax comes due later, and it comes due in the worst possible places.

It comes due in the meeting where someone challenges a number and you cannot defend the reasoning behind it, because there was no reasoning of yours to defend. It comes due when a task lands that is genuinely novel and there is no similar example in the training distribution for the model to lean on, and you discover you have not built the muscle for a blank page in eighteen months. It comes due at the interview, where the questions are specifically designed to test whether the work on your resume was yours. And it comes due in the promotion conversation, where the thing being evaluated is judgment, and judgment is the one thing that only accrues through repeated exposure to consequences.

It is worth noticing who pays the most. Early-career people pay the highest rate, because a larger share of their assignments are the assignments that were supposed to build them, and because they have the least ability to tell a load-bearing task from a disposable one. A senior person delegating a formatting pass loses nothing. A junior person delegating their first twenty analyses loses the twenty analyses.

The tax is not a punishment for using AI. It is the price of a specific pattern: using it on the work that was going to build you. Which means the rule is not "use less" — it is "know which is which."

Sort every task by one question: is this work meant to produce an output, or to produce me?

Spend freely. Formatting. Reformatting data between templates. Boilerplate you have written four hundred times. Transcript to summary. Renaming things consistently. Translating your rough draft into someone else's house style. First-pass literature scans where you already know how to evaluate what comes back. None of this builds anything, and doing it by hand is not virtue, it is friction.

Spend carefully, with the Explain-Back Rule enforced. Analysis you will be accountable for. Anything a client or executive will question. Work in your own specialty where your reputation is the deliverable. Here the model is a fine collaborator and a terrible ghostwriter — ask it to critique your draft rather than produce it, ask for the case against your conclusion, ask what you are missing.

Do it yourself. Skills you are actively trying to acquire. The first ten instances of any new task type. Anything where the reasoning is the product rather than the document. And the writing that has to genuinely be yours, because voice is not a formatting layer over content and readers can tell.

The one thing to never skip, in any bucket: checking factual claims. Fluent output has no tonal signal for invention, and the verification habit is separate from the understanding habit. The mechanical version of that check is in how to fact-check AI answers.

This is the reason Mochivia's lessons ask you to answer before they explain. Not to be difficult — because retrieval is the mechanism, and a lesson that lets you nod along has taught you the feeling of learning and nothing else. Every checkpoint is a small, deliberate refusal to let recognition stand in for recall.

The Explain-Back Rule is the same principle applied to your job instead of a lesson. You do not need us to run it. You need ninety seconds and a willingness to hear yourself fail it.

If you want a slightly stronger version, do it in front of a person. Explaining to a colleague who is allowed to say "wait, why" catches things an imaginary audience lets you slide past, and it has a useful side effect: the people who watch you reason out loud form a much more accurate picture of what you actually know than the people who only see your finished documents. That is worth something in a period where finished documents have stopped being evidence of anything.

The other half of using these tools well is stating what you want precisely enough to get it, which is a real skill and not the one most people think it is. That argument is in why prompt engineering is not the skill you need, and the full four-layer picture of durable AI competence is in how to learn AI skills that don't expire.

Two people start the same job on the same day and both use AI constantly.

One of them explains their work out loud before it ships, does the first ten instances of anything new by hand, and asks the model to attack their reasoning rather than supply it. The other ships faster all year.

For about eighteen months the second person looks better. Then a genuinely novel problem arrives, or a promotion panel asks a why question, or the model gets confidently wrong in a domain nobody on the team can check. At that point the difference between them is not effort or intelligence. It is that one of them kept doing the part that was building them.

Take the last thing you sent. Close the tab. Explain it out loud. Whatever happens in the next ninety seconds is the most honest performance review you will get this year.