Verifying AI output is a competence, and almost nobody teaches it

This is commentary. It comes out of reading course syllabuses while drafting D6, and the pattern was consistent enough to be worth stating plainly.

Almost every course we looked at teaches production: how to get output, how to get better output, how to get output faster. Almost none teaches verification: how to establish, before relying on it, whether the output is correct for the purpose. The syllabus ends at the point where the professional obligation begins.

Why the gap persists

Production is demonstrable and satisfying to teach. Verification is slow, produces no artefact the learner can show, and frequently concludes that the impressive output from the previous module should not be used. It is a harder sell and a less pleasant lesson.

It is also the competence that separates someone who can be trusted with an AI system from someone who cannot. Fluent, confident and wrong is the characteristic failure mode of these tools, and the whole of D6 L1 exists to establish that a person recognises it.

What a course could do about it cheaply

Give learners output that is wrong in ways they must detect, with criteria to apply, and mark them on the detection rather than the production. Require the evaluation record, not the verdict. Assess at least one statement about documenting a judgement so that another person can reproduce it, because a verification nobody can reproduce is an opinion.

None of that requires new tooling or a longer course. It requires deciding that the learner’s ability to distrust the output is part of what the credential asserts.