Verification is a professional act, not a chore

In most legal research instruction, verification is the last thing on the list. Students learn to find authority, assess its weight, and construct an argument, and then at the end they are told to cite-check their sources, confirm everything is still good law, and move on. The framing is tidy and logical: do the work, then check the work. And it is not wrong as far as it goes. Shepard's and KeyCite exist for a reason, and using them is genuinely important. The problem is that this framing positions verification as something that happens after the substantive work is done, a quality-control pass before submission rather than an integral part of the research itself. That distinction matters more than it might seem, and it matters most when AI is in the workflow.

The argument this note is making goes past the question of when or how carefully students verify. The underlying problem is that they understand verification as the wrong kind of task. When a student opens a case to confirm it says what the AI said it says, they are reading the case. When they confirm that a statute has not been amended since the AI's training data was gathered, they are establishing what the law currently is. When they compare the proposition in the brief to the actual holding in the source, they are doing legal analysis. Verification, understood correctly, is the substantive work itself, and a student who internalizes that will approach AI output differently than one who treats verification as cleanup. That difference shows up most clearly when the errors are the invisible kind: the real citation with the wrong holding, the standard applied in the wrong jurisdiction, the plausible-sounding fact the AI inserted because no one told it the fact was wrong.

The reframe is easy to say and genuinely difficult to make stick. Students arrive with years of prior instruction that positioned verification as the administrative tail end of the research process, something to complete before they could consider the assignment done. One lecture does not undo that. What the course does structurally to reinforce the reframe rather than just assert it is worth naming explicitly, because instructors adapting these materials will need to make similar choices. The Red Line Challenge is designed around the insight that students learn verification habits by doing verification badly and seeing the consequences. In that exercise, students are given an AI-generated brief seeded with multiple error types, including fabricated citations, citation drift, doctrinal distortion, assumed facts, and tone mismatches, and asked to flag everything they find without using any tools. The results are instructive in the most concrete possible way: students tend to catch the obvious errors, a case name that turns up nothing in Westlaw, a statute that does not exist, and consistently miss the harder ones, the case that is real, cited correctly, and says nothing like what the brief claims it says. That gap between what students found and what was actually there is the pedagogical content of the exercise. The reframe lands when students experience it, not when they hear it.

Traditional verification in legal research was already necessary and already undervalued as a teaching priority. AI makes both the necessity and the undervaluation worse, and the reason is fluency. A hallucinated citation reads like a real citation. A case summary that has no relationship to what the case actually holds is written in the same confident, well-structured prose as a summary that is accurate. The Feldman matter, which the course uses as a case study, illustrates this with particular clarity: the attorney submitted briefs with fabricated citations across multiple filings, was warned by the court, submitted a response to the order to show cause that itself contained a hallucinated citation, and then filed yet another brief with additional fabrications while a sanctions hearing was pending. By his own account, he did not read the cases. He fed citations through AI programs and assumed the output was reliable. What makes the case useful pedagogically is not only the magnitude of the failure but the mechanism: AI produces outputs that look like the product of professional legal work, and that appearance creates a specific kind of risk that traditional research did not present in the same way. When a case summary looks authoritative, the instinct to verify it is harder to activate. Students need to be trained to activate it anyway, which means they need to understand that the fluency of the output is the risk, not a sign that the output is trustworthy.

Whether and how to grade verification is a genuine instructional decision with real tradeoffs, and instructors adapting these materials will land in different places depending on their course structure and institutional context. The Red Line Challenge is ungraded in this course, and that is a deliberate choice. An ungraded exercise surfaces what students actually do rather than what they do when they know they are being assessed, and the class debrief of the results is where the learning happens. Students who found two out of five fabricated citations learn something important about their own verification habits when they see the full error count. Students who ran a database search and stopped there learn something different when they see that the tools caught all five fabricated citations but zero assumed facts, zero doctrinal distortions, and zero tone mismatches. The gap between tool performance and human performance on the invisible errors is the most important data point in the exercise, and it only lands if students have first worked through it themselves. Instructors who choose to grade verification work should think carefully about what they are signaling. Graded verification tells students that checking boxes counts; ungraded verification followed by a thorough debrief tells them that developing the underlying instinct is the point. Both approaches can work, but the question worth asking first is which one the course is actually trying to build.

Feldman is used throughout this note because the procedural record is unusually complete and the error escalation across multiple filings makes the verification failure visible in stages. Any case involving AI-generated hallucinated material can serve the same pedagogical function, however, and instructors should feel free to substitute a more recent example or one that better matches their students' practice context. The landscape of sanctions decisions is developing quickly, and a case decided closer to the semester may land with more immediacy than one students have already encountered in other courses.