Most organizations have a plan for AI-generated training. Someone reviews it before it goes live. A manager signs off. Box checked.
Here's what that plan misses: a human who can't recognize a badly designed learning experience can't catch one that AI produced either. And right now, a lot of organizations are approving training that looks professional, reads clearly, and quietly produces no learning at all.
The people most excited about AI outputs are often the least equipped to question them
In 2025, researchers tested AI receptivity across six studies with a nationally representative U.S. sample. The finding was consistent every time:
Lower AI conceptual knowledge predicted higher willingness to use and trust AI outputs.
Why? People with less understanding of how AI works were more likely to experience it as magical.
Magic doesn't get questioned.
Nielsen Norman Group found the same pattern playing out in real usage. They identified a specific user type: the naive power user. Fluent at prompting. Comfortable with the tools. Largely uncritical of what comes back.
These aren't careless people. They're confident ones. And confidence without domain knowledge is exactly where errors get approved.
What an expert actually catches
Here's what that looks like in practice.
AI produces a module where all the practice questions cluster at the end — killing the spacing effect before it has a chance to work. It goes to review. It gets approved. Nobody flags it, because nothing looks wrong.
It uses humor. Genuinely funny humor. A character who's catastrophically overconfident, a scenario that makes people laugh out loud. The problem: the joke has nothing to do with the concept it follows. It encodes separately. It retrieves separately. It's decoration dressed as design. That distinction requires knowing why connected humor works — and what happens when the connection isn't there.
It front-loads every piece of information a learner needs before giving them anything to do with it. Maximum cognitive load at exactly the moment working memory is most constrained.
Looks thorough. Functions badly.
The module has objectives. A summary. A completion button.
What it doesn't have is a well-versed learning professional vetting it.
The reframe & the honest implication
"Human in the loop" catches factual errors. Wrong date, bad link, compliance gap.
"Expert in the loop" catches structural errors. The design decisions that determine whether learning actually happens.
Not the same job. And here's what that distinction actually requires: you have to be the expert. Not in title. In knowledge.
A lot of L&D practice runs on production skills, tool fluency, and institutional familiarity. Not on a working understanding of how memory consolidates. Not on why spacing practice matters, or what cognitive load theory says about sequencing. Those aren't required for most completions to go green. They're not in most job descriptions.
But they're exactly what AI can't replicate in its outputs. And exactly what a non-expert reviewer won't catch when they're absent.
Think about marketing. A legal reviewer clears copy for compliance. Only a marketer who understands persuasion, audience psychology, and the mechanics of attention can determine whether the copy will actually work. Legal can approve it. That doesn't make it good.
Your domain knowledge works the same way. It can't be substituted with oversight. It has to be built. Then applied.
The practitioners who will matter most in an AI-assisted L&D environment aren't the fastest tool users. They're the ones who have developed enough knowledge to know when something is wrong and why.
