Evaluate before you rely

I did not teach this way the first time I ran this course. In that first year, my instinct was to give students as much access as possible, let them try anything, form their own impressions, develop their own workflows. It was an instinct toward openness, and at the time, it felt like the right one. What I did not anticipate was how much the tools themselves would shape what students thought evaluation even meant. In 2023, the outputs were uneven enough that students did not need a framework to identify a bad result. They just knew. The response was either cautious optimism or outright dismissal, and neither of those postures produced the kind of critical engagement I was looking for. Students who thought the tools were impressive did not interrogate the outputs. Students who thought the tools were disappointing stopped engaging with them at all.

The problem was not that I was teaching evaluation too late. The problem was that I was not teaching evaluation at all, I was hoping it would emerge from exposure, and that is not how judgment develops.

By the second year, something had shifted on the student side that made this even more apparent. Nearly every student arrived having already used AI tools for legal work, for coursework, for everyday research tasks. The “this is a waste of time” reaction was largely gone. What replaced it was something harder to work with: comfort without competence. Students had formed impressions of what good AI output looked like, and those impressions were based on fluency, not accuracy. A well-written hallucination read as a good result. A hedged but correct answer read as weak.

This is why the course now leads with evaluation exercises rather than tool introductions. The goal is not to make students skeptical before they have anything to be skeptical of, it is to give them a framework before fluency hardens into habit. Asking students to think through what kinds of mistakes a tool is likely to make, and how to recognize those mistakes in an output, produces something different than asking them to rate whether a result was helpful. It produces a posture. And that posture travels with them into every tool they use, not just the one they happened to try first.

The Week 1 exercise is not, on its surface, an evaluation exercise. Students are asked to input a prompt into any tool, we have already established in the first session that we will use any tool throughout the course, and for this exercise, they just need to use something. We discuss what research is, why I define it the way I do, and what the course is actually training them to do. The prompt exercise is a test of their instincts at baseline: what they reach for, how they construct a question, what they assume the tool needs from them.

The evaluation frame enters in session two, when we pull that exercise forward and look at it differently. Now the question is not “what did you get” but “what could have gone wrong here, and would you have caught it.” The exercises are designed to produce outputs that look plausible but fail in ways that require legal knowledge to identify. That friction is not incidental, it is the whole point. A student who has experienced that friction once approaches every subsequent output with a different question running in the background.

When tool fluency comes later in the course, students learn it in a context where they already know what failure looks like. That sequence matters because it means confidence, when it develops, is earned. Students who become fluent after developing an evaluative framework tend to be more accurate about what they can trust their tools to do and what they cannot. Students who become fluent first have to be walked backward through that reckoning, and it is a harder journey, it requires them to question something they have already built a workflow around.

The goal is not a classroom full of skeptics. Skepticism without skill is just another form of disengagement. The goal is students who can use these tools at a high level precisely because they understand the failure modes, not in spite of it.

The resistance I expected in session one, students frustrated that we were not just diving into the tools, largely did not materialize, and I think the structure explains why. The first class is oriented around what research is and why it matters, and the initial exercise asks students to use a tool without any particular constraints. It does not feel like deprivation. The evaluative turn happens in session two, by which point students have something concrete to evaluate rather than an abstract question about methodology.

What I watch for instead is the student who, when asked what could have gone wrong with an output, immediately reaches for “the AI hallucinated” as a complete answer. That response tells me the student has learned the vocabulary without doing the underlying work. Hallucination is a category, not an analysis. The exercise is designed to push past that, to ask not just whether something went wrong, but where, why, and whether the error was the kind a careful reader would catch. That is the judgment the course is building, and it starts here.