Generative AI has become as routine in legal practice as the yellow highlighter. CoCounsel summarizes productions. Westlaw AI surfaces cases. ChatGPT drafts the first pass of a memo before a junior associate ever opens a blank document. The technology is now part of how the job actually gets done.
But the story law schools are telling about AI, if they’re telling one at all, is lagging behind the story unfolding in legal practice. As a recent graduate now working at the intersection of legal tech and legal education, what I keep hearing from young lawyers, judicial clerks, judges, and other recent graduates is a profession improvising. New lawyers are learning powerful systems on the job, often without clear institutional guidance, shared professional norms, or confidence in their own ability to supervise the output.
That gap matters because legal education still shapes how lawyers understand professional competence. Law school does more than teach doctrine. It builds the foundational habits of legal practice: rigorous research, disciplined writing, and the kind of careful evaluation of authority that doesn’t happen by accident. AI is altering all of those habits at once, and yet most students are still encountering AI tools informally, through internships, summer associate positions, online experimentation, or whatever they try alone before a memo is due.
In many workplaces, the absence of a clear AI policy hasn’t slowed adoption. It’s just driven it underground. Associates open ChatGPT in a browser tab or on a personal device. Clerks quietly test research tools against traditional methods. The result is a legal profession drifting into a shadow-IT model of professional formation: informal trial-and-error with no shared supervision standards, happening at the exact point in a lawyer’s development when judgment is supposed to be forming.
That institutional ambiguity is especially visible from inside legal education. Some students graduate having spent serious time with AI drafting and research tools. Others leave with little more than a warning about hallucinated cases. Even within the same law school, students and recent graduates describe radically different levels of exposure depending on which professors they happened to take, which clinics they joined, or whether they pursued the subject independently. Students are entering practice having used AI without understanding how to use it well.
The deeper problem isn’t that students are using AI. It’s that AI is reorganizing the order in which professional judgment develops.
Legal education and early practice have historically assumed that competence develops sequentially: first research, then drafting, then judgment. A junior associate read cases, produced a rough memo, received edits, missed nuances, and gradually built the instincts that allow a senior lawyer to recognize when an argument is structurally weak or a citation is being stretched. Generative AI scrambles that sequence. A first-year lawyer can now produce polished work product before developing the professional instincts necessary to evaluate it.
That inversion is what the young lawyers I spoke with kept circling back to, even when they didn’t describe it in those terms. One associate told me she uses Harvey constantly and still reads every sentence line by line. She isn’t worried about obvious hallucinations. She’s worried about the ones that sound right. A judicial clerk described an office where staff scan briefs for hallucinated authority because fabricated citations have become common enough to anticipate. Another first-year admitted AI allows him to produce work “above my class year,” while worrying he lacks the judgment to know whether the analysis underneath is sound.
Generative models are very good at producing analysis that looks professionally competent. The difficulty is that junior lawyers often lack the experience necessary to distinguish polished reasoning from sound reasoning. That distinction used to develop in the same process that produced the work product itself. Now the work product arrives first, and the evaluative instinct has to catch up to it in real time.
Drafting flaws are usually visible. Supervisory failures are subtler. They show up as an overconfident synthesis of authority, a missing qualification, a persuasive paragraph that quietly sidesteps the hardest legal issue, or a citation that technically supports a proposition while distorting what the case actually said. Catching them requires the kind of instinct that traditionally developed over years of producing legal analysis under supervision and watching senior lawyers catch things you’d missed.
That changes what junior legal work actually is. The cognitive center of the job is shifting from production to supervision: checking citations, pushing back on reasoning, restructuring arguments, comparing summaries against what the source actually says. Supervision is harder than drafting, and it draws on instincts a first-year hasn’t had time to develop. A junior associate hasn’t been at this long enough to feel when something’s off, when an argument quietly sidesteps the hard question, when a paragraph reads well but isn’t actually saying much. That kind of recognition takes years. AI is asking new lawyers to perform it on day one.
The optimistic case is real and worth taking seriously. AI may give junior lawyers earlier exposure to higher-order analytical work instead of trapping them in rote production. Reviewing generated output repeatedly may, over time, sharpen evaluative instincts faster than the traditional path would have. But the optimistic case still has to reckon with what it’s accelerating past. Effective supervision itself requires judgment, and legal education has long assumed that judgment develops after extended experience producing legal analysis independently. The foundation AI is letting junior lawyers skip past may be the same foundation that makes the skipping safe.
Legal education has traditionally emphasized the production of legal analysis from primary sources upward. Increasingly, young lawyers are instead being asked to supervise machine-generated analysis under time pressure.
Students need structured practice doing exactly that work. They need experience identifying when a polished answer is subtly wrong, tracing AI-generated claims back to source authority, and testing the limits of systems that produce persuasive language without reliable reasoning underneath it. They need training in judgment itself: when to trust the tool, when to slow down, and when generation is doing the easy work in place of harder thinking.
That exposure has to include both the legal-specific platforms students will encounter in firm and clerkship settings and the frontier general-purpose models that are already shaping how lawyers draft and reason. Law schools devote enormous effort to teaching Bluebook citation and traditional research databases because no one expects students to learn them by osmosis. The same standard should apply to AI.
Prompting deserves a particular place in that curriculum, because it draws on skills law schools already teach. Defining constraints, giving the model the right factual context, writing instructions that don’t leave room for guesswork: these are the same disciplines that underlie a well-framed research question or a clearly drafted memo. The quality of AI output depends heavily on the quality of the query, much as the quality of legal research depends on how a problem is framed before a search is run. Law schools already accept that research platforms require instruction. AI tools belong in the same category.
The lawyers most comfortable with these systems are often the least utopian about them. Extended exposure tends to produce caution. One associate at a firm heavily invested in AI described herself, only half-jokingly, as “a semi-hater” because effective prompting and careful review can take as long as doing the work manually. That kind of judgment, knowing when the tool helps, when it slows you down, and when its fluency is masking a thin underlying argument, is itself a form of professional formation. It develops through repeated, supervised exposure, which is precisely what most students aren’t getting.
Legal educators have absorbed every previous shift in how lawyers work, electronic research, e-discovery, clinical practice, the formalization of legal writing instruction, by treating it as a curricular question rather than a technological one. Generative AI is doing something different. It is beginning to change how lawyers think, and the developmental sequence through which they learn to think like lawyers at all. The shift is already here. The question for legal education is whether to shape it or trail it.