Teaching a course with AI Twins: what we learned

When we ran a law school course with AI knowledge twins available to students alongside the instructor, the interesting result was not that it worked. It was what students used them for, which was not what anyone predicted.

What we expected

The assumption going in was that students would use an AI Twin to shortcut the reading. Ask it what the case said, skip the case.

Some did, briefly. It is a bad strategy and the course structure exposed it quickly, because the assessment asked for reasoning rather than recall.

What actually happened

The dominant use was rehearsal. Students used the AI Twin to check whether they had understood something before saying it out loud in a seminar where being wrong is expensive.

This is the thing a busy instructor genuinely cannot provide. There is one of them and thirty of the students, and the question a student most needs to ask is the one they are embarrassed to ask in front of peers. An AI Twin grounded in the course material answers that question at 11pm without judgement, forty times if necessary.

The second pattern was navigation. Where in this material does the argument about X appear? Students used the AI Twin as a way into a corpus rather than a substitute for reading it.

What broke

Three things, all instructive.

• ·Material gaps became visible immediately. Anything not in the uploaded corpus produced either a refusal or, worse, a generic answer from general knowledge. The fix is curation, not cleverness: put the actual material in.
• ·Students over-trusted fluent answers. A confident paragraph reads as authoritative regardless of whether it is grounded. The course had to teach the habit of asking the AI Twin where an answer came from.
• ·The AI Twin was better at settled doctrine than at contested questions, which is exactly backwards from what a law course is interested in. Contested questions need judgement, and judgement is the part that stays with the instructor.

The conclusion that matters

An AI Twin did not reduce the amount of teaching. It changed which teaching happened.

Time previously spent on the twentieth explanation of a foundational concept moved to argument, edge cases and application. That is the part of a course only a person can run, and the part that was always getting squeezed.

What this generalises to

Any body of knowledge taught repeatedly to cohorts has the same shape. Corporate training. Onboarding. Professional certification. Anywhere the same foundational questions arrive every cycle from a new group of people who are new to them even though you are not.

The instructor keeps the judgement work. The AI Twin absorbs the repetition. Neither one replaces the other, and the value comes entirely from getting the split right.

This piece revisits an earlier essay from our Medium archive: Education in the age of AI.

See how educators use AI Twins.