Knowing What You Handed Over

The debate has been stuck between two poses. One treats offloading as a leak we should resist. The other treats delegation as pure upside, the more you hand over the better. Both miss the same point. Dividing cognitive work between a person and a machine is a skill, and like any skill it can be done well or badly. The quality of the division is what matters, not the quantity of the offloading.

To talk about quality, here are the two terms.

Cognitive calibration

Cognitive calibration is the accuracy of your judgment about who should do what, made as you go, based on a realistic read of your own capability and the AI’s.

The word is precise. A calibrated instrument reports values that match reality. A calibrated person holds beliefs about their own competence, and the tool’s, that match what each can actually do. Miscalibration runs both ways. Overrate yourself and you refuse to hand over things the model does better. Overrate the model and you give away judgment you were better placed to make.

This failure is already documented, and the finding is sharper than you might expect. When humans and AI team up on a classification task, the pair beats the AI alone only when the AI decides how to split the work. When the human decides, the gains vanish. The reason is that people cannot accurately judge their own ability, so they delegate badly. The researchers call the missing ingredient metaknowledge, and they make a point that should stop anyone designing these systems: lacking it is an unconscious trait. You do not feel yourself misjudging the split. That is exactly why calibration has to be named and trained rather than assumed.

Calibration is forward-looking. It is the question you ask at the start, which parts are mine to hold and which are better given away, and the question you keep asking as the work moves and the picture changes.

Cognitive accountability

Cognitive accountability is the capacity to give a truthful account, afterwards, of how the work was actually produced: what you contributed, what the AI contributed, where they combined, and where you genuinely cannot tell.

If calibration looks forward, accountability looks back. It is not about whether the division was good. It is about whether you can honestly say what the division was. The two come apart more than you would think. You can calibrate well and still lose the seams, ending with strong work whose provenance you cannot reconstruct. You can calibrate badly but stay aware of it, which at least leaves you able to correct.

Accountability earns its place because it matters for reasons that have little to do with each other.

It is epistemic. If you cannot say which parts you reasoned to and which arrived fully formed from a model, you do not know how far to trust the result. The confidence you feel is borrowed, and borrowed confidence is the dangerous kind.

It is practical. When something turns out wrong, the person who knows which step was theirs and which was the machine’s can find the fault and fix it. The person who cannot has to discard the whole thing, because they have no map of where the error might live.

It is professional, and this is where it bites hardest. In medicine, in research, in any field where a claim has to be stood behind, the provenance of the reasoning is not a nicety. A clinician who cannot say which part of a recommendation was their own judgment and which came from a model has a problem that no amount of correctness in the output resolves.

Why both, not one

The terms need each other. Calibration without accountability gives you someone who divides work sensibly but loses the thread, so they cannot learn from the work or defend it. Accountability without calibration gives you someone who can describe a badly run process in perfect detail, which is honest but not much help. Together they describe the whole practice: divide the work well, then be able to account for the division you made.

There is a shortcut worth refusing. It is tempting to collapse both into “stay in the loop,” the phrase everyone reaches for. It is not wrong, just empty, because it never says what being in the loop consists of. These two terms are an attempt to say it. Anything less is a feeling of involvement, and a feeling of involvement is precisely what a fluent model is best at manufacturing.

What it changes

Name these and the effort moves somewhere more useful. If the risk were simply offloading, the response would be to offload less, which is neither realistic nor desirable. If the risk is poor calibration and weak accountability, the response is to get better at two things you can actually practise, neither of which means using AI any less.

For anyone teaching people to work with these tools, that is the better target. The question stops being whether students use AI and becomes whether they can judge the division and account for it. Both are assessable. Ask someone to defend why they split a task the way they did. Ask them to walk back through a finished piece and mark what came from where. Someone who can do both is working well with AI in a way “offloading” could never capture. Someone who can do neither is exposed, no matter how little they appear to lean on the machine.

The terms are deliberately plain. Calibration, the accuracy of the split. Accountability, the truthfulness of the account. I would rather two plain words that get used in a seminar or a design review than one clever one that stays on the page. The whole point is to pick them up and put them to work.

Adrian Cowell is Innovation Lead at Imperial College London’s Faculty of Medicine and co-founder of World1-1 Studio. He works at the intersection of education, emerging technology, and games.