Last year I made a small change to my graduate NLP seminar. Instead of forbidding AI assistants, I required students to use them — and then to write a critique of what the model got wrong.
What worked
Students became sharper readers. Catching a model's confident mistakes turned out to be excellent training for catching their own. Arabic-speaking students, in particular, found rich material in how models mishandled dialect and cultural context.
The most valuable skill I can teach now isn't producing text. It's judging it.
What didn't
Some students leaned on the tools for first drafts and struggled to find their own voice. Weaker writers sometimes produced polished text they couldn't explain. Assessment had to shift toward conversation: short oral defenses of written work.
What I'm keeping
- Critique assignments that make model errors the object of study.
- Oral check-ins for major written work.
- Open discussion of when using AI is — and isn't — appropriate.
I don't think there's a final answer here. But treating these tools as something to think with, rather than something to hide from, has made for a more honest classroom.
