Teaching Students to Fly With AI, Not Around It

There has been a lot of news lately about law schools banning the use of AI — the University of Chicago is one of the more extreme cases, prohibiting computer use for the entire first year of law school. I understand the impulse: AI has made prior learning methods like take-home exams and homework essentially useless as measures of competence. But the question I keep coming back to is this: how do we design a system that doesn’t just exclude AI, but lets students and AI co-exist and co-evolve? Like it or not, AI is here to stay. You can ban it as hard as you want, but students will use it anyway. The right approach isn’t to ban AI as much as possible — it’s to teach students how to master it and incorporate it correctly into their workflow.

Lessons From the Cockpit

This brings me back to a talk I keep thinking about, given at the Bloomberg Law Symposium by Alistair Wye, Director of Innovation & AI at White & Case. His presentation drew a parallel between autopilot in aviation and AI in legal practice, and it centered on four failure modes:

• Skill decay — “Juniors who have never drafted from a blank page. Associates who never cite-checked unaided. The skill does not announce its departure.”
• Complacency — “‘I checked the output’ becomes the legal equivalent of ‘the autopilot has it.’ Checking requires a standard to check against, supported by substantive expertise.”
• Emergency dependency — “The system gets a subtle answer wrong. Who on the team has the competence to catch it? If the answer is ‘the partner,’ we have a succession problem.”
• Mode confusion — “Retrieval, reasoning, and generation produce outputs that look identical. Lawyers treat hallucinations and citations the same way — on the page, they are; but in practice the underlying failure modes may be different.”

All four are real. But I don’t think they’re equally concerning.

Skill Decay and Emergency Dependency: Less Alarming Than They Sound

Take skill decay. I wouldn’t call it “decay” so much as a shift — we have different tools now, so we need a different set of skills. (I’ll spare you the calculator analogy; I’ve used it too many times already.) But consider: with the sheer number of free templates available, is anyone actually drafting from a blank page anymore? For most contract types, every firm already has a template it modifies as needed, and those templates aren’t going away. Even without AI, the skill juniors need was never really “draft from nothing” — it was always “modify from a template intelligently.” AI doesn’t erase that skill; it just changes what modifying well looks like.

Emergency dependency, meanwhile, I’d frame less as a crisis and more as the new differentiator for who makes partner. Yes, the legal AI landscape is chaotic right now — adoption has happened almost overnight, with no gradual runway to ensure backward compatibility. But the older generation of partners isn’t retiring yet, and they remain the competent backstop on the team. The associates who can bridge the gap between the two generations of practice are the ones positioned to become the next generation of partners. That likely means a different set of criteria for partnership. A few years ago, making partner was largely a function of billable hours — a long, high-volume grind. In the AI era, it may become less about how many hours and more about how effectively they’re used: how quickly an associate can deploy the tools at hand to solve a problem, including catching the subtly wrong answer.

Complacency and Mode Confusion: The Ones That Actually Matter

The other two lessons deserve more weight, because they get at where legal education — and arguably all education in the AI era — needs to go.

Start with complacency. There’s nothing inherently wrong with being complacent, if you know the tool you’re using is reliable and you’re using it the right way. I made this case in more detail in a previous post. The core idea: know what a tool is designed for, and use it only for that. If you know a model hallucinates, don’t take its output at face value — verify it. Verification is time-consuming, so the goal should be to design workflows that make checking modular and fast. Once you’re confident you’re using a tool exactly as intended, and you’ve stress-tested where it can go wrong, complacency stops being a liability and becomes earned trust.

Mode confusion is really the same problem underneath. Both come down to genuinely understanding the tool you’re using — not just its outputs, but its mechanics and its blind spots.

What Law Schools Should Actually Teach

So what should legal education do to help students build this understanding?

I keep coming back to this Berkeley Law piece as a model for how to approach it. Students shouldn’t be expected to intuit how (or how not) to use AI — not just because many come from non-technical backgrounds, but because the landscape itself is genuinely hard to navigate even for experts. There are multiple foundation model families, each with several models built around different goals, and on top of that a growing set of legal-specific tools like Harvey and Legora, each with their own strengths.

A curriculum that teaches students what each model family is actually built for — and has them compare results on tasks the model was designed for versus tasks it wasn’t — would make the failure modes tangible rather than theoretical. If you think of AI as a collaborator, the lesson is simple: you don’t hand a collaborator work outside their specialty and hope for the best.