What to Check Before an AI Lesson Reaches Students

AI-generated lessons fail in predictable places, which makes them easier to check than you’d think.

You already read everything before it goes in front of students. None of this is a new job. It’s a shorter version of one you’re doing anyway, aimed at where generated drafts actually go wrong.

Six things, roughly in the order they cost you if you miss them. None of this is specific to CARL. It works on anything a model hands you.

Check the facts that would be embarrassing to get wrong

Not everything. The handful a student could look up in twenty seconds.

Dates, names, numbers, the plot of the book, the direction a chemical reaction goes. Generated text is fluent in a way that makes wrong things read exactly like right things, which is precisely the problem. Fluency isn’t a signal. If a claim would make you wince coming back at you from the third row, verify that one.

Check that the sources are real

Then check whether they say what the lesson claims they say.

Generated tools can invent citations, mix up titles, or attach a real source to a claim it doesn’t support. If there’s a link, open it. If there’s a title without a link, search it. If it comes from an organization you don’t recognize, start there.

A fake source in a lesson about evaluating sources is a particularly bad day, but the rule applies everywhere.

You can also put the AI to work on this. Ask it to list its sources with links, then ask it to check its own claims against them. It won’t catch everything. It can be confidently wrong about its own confidence. But it turns a vague worry into a list you can click through.

Read who is in the examples

Names, families, roles, and who is pictured doing what.

Models lean on defaults. Defaults are narrow unless something pushes back. Is the scientist always the same kind of person? Do the families all have the same shape? Has somebody’s background been turned into a costume instead of just being part of who they are?

The other half is context you have and the model doesn’t. Whether an example lands in your community. Whether a topic sits differently for a student in your room. Whether the framing assumes experiences your students don’t share.

Make sure the supports attach to something real

Generated supports are often reasonable in the abstract and useless in the specific.

“Provide sentence frames” isn’t yet a support. A sentence frame for the actual task is. Read each accommodation against the steps of the activity. Does it name a real barrier in this task? Does it give the student something concrete to do about that barrier?

If a support could be pasted into any lesson on any topic, it probably needs another pass.

Read the assessment against the activity

Not only against the outcome. Against what students were actually asked to do.

This one hides best. Both pieces look fine alone. The rubric measures analysis; the activity asked students to summarise. The success criteria describe a skill that never came up. Built separately, a plan and its assessment drift, and the drift is invisible until you’re marking.

Read it as the day you’ll actually teach it

The last check is the one only you can run.

Timings, on the day of the assembly. Group work, with that combination of students. The discussion activity, in the week it would actually land. A lesson can be structurally sound and still be wrong for the room.

No amount of context catches all of that because the thing being judged isn’t only the lesson. It’s the fit.