This summer’s associate classes learned something new in their first week: how a large language model works. Law.com reports that AI skills are now a standard target of summer training across the Am Law 100, the maturing of a practice Bloomberg Law documented when K&L Gates, Orrick, and Dechert first built generative-AI training into their summer programs, with K&L Gates and Orrick bringing in an outside provider, AltaClaro, to teach prompt engineering. Ropes & Gray (my firm) made generative AI a core curriculum requirement for its more than 170 summer associates in 2025, and in 2026 paired summer associates with first-year associates in the TrAIlblazer Cup to build working AI solutions for real legal workflows on billable time. On paper, this instruction should be redundant. The ABA’s AI and Legal Education survey, summarized in the AI Task Force’s Year 2 report last December, found that 55 percent of law schools offer classes dedicated to AI and 83 percent offer curricular opportunities to learn legal AI tools. But ask the students. In Thomson Reuters’ 2026 Law Student Pulse Survey of more than 1,800 law students, 32 percent said their school does not provide the AI skills their careers will require, 48 percent said AI policy varies professor by professor, and the majority who use AI several times a week reported teaching themselves. The coverage the schools report is elective, uneven, and easy to graduate around, so the firms are re-teaching the foundations to everyone.
That arrangement is backwards, and the way to fix it is to divide the curriculum by shelf life. Law schools should own the durable skills: how models produce text and fail, prompting and context fundamentals, verification, and the professional-responsibility rules that govern all of it. Firms should own the perishable ones: their own tool stack and the workflows of their practice groups. And firms hold the instrument that would enforce the division, because they can evaluate the durable skills at hiring, the same way they have always evaluated research, analytical skills, writing, and whatever else they value in junior associates.
The division of training
What belongs to the schools is everything that will still be true when the current tools are gone: how a model generates text and why it fabricates authority when a prompt provides insufficient grounding, why output degrades over a long context, the principles of grounding, scoping, and iteration that follow from those failure modes rather than from any vendor’s interface, and the verification discipline of treating a generated proposition as unproven until an independent source proves it. The most durable layer is professional responsibility, which the schools are already accredited to teach: technological competence under Comment 8 to Model Rule 1.1 and ABA Formal Opinion 512, the confidentiality analysis for feeding client information to a model, and the disclosure obligations that vary by court.
What belongs to the firms is everything the schools could not teach if they tried. Law schools have never taught the workflows of practice: most students graduate without ever having run a due-diligence review or marked up a credit agreement, the daily work of the corporate associates many of them will be within months, and nobody faults the schools for it, because that training exists only where the deals are. AI has made the workflows no more teachable from a classroom. No school can teach a deal team’s AI-assisted diligence workflow, the firm’s retrieval setup over its own precedent bank, or the way a particular client’s guidelines constrain which tools may touch which matters. That training is inseparable from practice, and it perishes fast even inside the firm. The tool list from Bloomberg Law’s earliest summer-associate reporting makes the point on its own: Casetext’s CoCounsel, one of the named platforms, has since been swallowed into Thomson Reuters’ product line, and the interfaces those first summer classes trained on have been versioned out from under them. A law school that had built a course around any of those products would have graduated students trained for software that no longer exists. The one that taught why grounded prompts suppress fabrication graduated students who can use whatever replaced it.
Some schools are demonstrating that the durable curriculum fits inside a law degree. Texas wrote “the signal importance of output verification” into its learning objectives. Case Western now requires every first-year student to earn a legal-AI certification, and UC Law SF (formerly UC Hastings) will require every JD student, starting with the class of 2029, to complete its AI-enabled lawyering lab, which pairs hands-on AI work with privilege, conflicts, and professional responsibility. Chicago sequenced the machine after the foundation but built supervised use into the required first-year writing course. None of these programs is tied to a particular product, and none will need to be rebuilt when the vendor list turns over.
What remediation costs
When a school defaults on the durable curriculum, the firm can remediate, and this summer shows firms doing exactly that. However, the costs of this arrangement are hidden in what the remediation displaces. A summer program is a fixed number of weeks; an associate’s first year has a fixed number of training hours that survive contact with billable pressure. Every session spent on what a context window is, or why models fabricate citations, is a session not spent on the training only the firm can give: how this practice group runs a deal, what this client’s guidelines permit. The firm is spending its scarce, practice-specific training capacity teaching what a required 1L course could have taught better, with assessment, to everyone at once. The price is no longer abstract, either. A firm that pays for AI learning in billable-hour credit, as Ropes & Gray now does at up to 400 hours per first-year, has put a dollar figure on those hours, and every credited hour that goes to model fundamentals purchases something a required 1L course had already been paid to deliver.
The graduate pays the longer-running share of the cost. A junior who arrives fluent in the durable skills starts compounding practice-specific judgment from the first assignment; the junior who spends the first months learning fundamentals starts that accumulation later, and the gap does not close on its own, because the work itself is shifting underneath both of them. As more of a practice becomes AI-enabled, the associate who can already scope, ground, and verify gets staffed on that work, gets supervised doing it, and builds the record of watched verifications that I have argued is the new apprenticeship. The associate still learning what the model can be trusted with becomes the one partners hesitate to staff, on the very work that is becoming most of the work. Michele DeStefano’s warning about summer associates, that without the substantive background you cannot discern whether the model’s answer is good, is usually read as a caution about doctrine. It cuts the other way too: without understanding the machine, the associate cannot discern which outputs even need the doctrine applied to them.
The self-taught cohort fares worst, and it is the largest one. The students filling the schools’ coverage gap on their own, the majority in the Thomson Reuters survey, are acquiring exactly the habits an unsupervised workflow produces: accept what reads well, skip the verification pass nobody taught, carry the practice into a job where the sanctions ledger is public and growing. The students themselves are not confused about the stakes. Seventy-two percent called AI literacy essential to their careers, and 74 percent said they worry that over-reliance will erode their own legal skills, which is a more sophisticated position than a curriculum that treats AI as an upper-level elective.
Evaluate the competency, not the tool
Law schools respond to what employers reward. The hiring market has started to reward AI experience, with lateral hiring of AI-experienced associates up 106 percent year over year, and Law.com reporting that AI literacy now functions as a differentiator in Big Law recruiting while visible aversion to the tools has begun to disqualify. The signal is arriving, but it can select for the wrong thing.
The wrong version asks for tool experience, and some of the recruiting conversation is drifting that way. A 2L’s tool list is mostly a fact about their school: which vendors gave it licenses, which clinic had the budget, which professor ran the pilot. Screening on it reproduces the school’s resources in the candidate pool and rewards exposure over understanding, and the understanding is the only part that survives. The firm interviewing for Harvey familiarity in August 2026 is interviewing for knowledge that its own procurement decisions could make obsolete before the candidate starts.
The right version evaluates the durable skills directly, and they are evaluable in an interview in a way tool fluency never was. Ask a candidate how they would verify a research memo a model drafted, and listen for whether verification means reading the cited case or re-asking the model. Ask when they would disclose AI use to a court or a client, and what they would refuse to delegate at all. Better still, do what writing samples have always done for writing: have the candidate walk through a piece of AI-assisted work, prompts and verification trail included, and grade the process. A candidate who can show a grounded prompt, name what they checked and against what, and say where they overrode the model has demonstrated the supervisory competence firms claim to want, whatever software they demonstrated it on. Transparency belongs on the same list, because a candidate’s willingness to show the trail predicts an associate’s willingness to keep one, and the record of AI use is becoming part of the work product itself.
Firms that evaluate this way get the associates they need, and they also get something cheaper than any training program: leverage over the schools. Accreditation moves slowly, and the ABA’s Task Force can urge curricular change only in the voice of recommendation. Hiring criteria move faster. The schools that overhauled legal writing when firms complained about it, and built clinics when firms rewarded practice-readiness, will build the durable AI curriculum the season on-campus interviewers start asking every candidate how they verify. A firm that both teaches remedial fundamentals and declines to ask for them at hiring is subsidizing the gap it complains about.
The division of labor is simple to state: schools teach what outlasts the tools, employers teach the tools, and the market tests whether the schools did their part. AI has not changed that structure, but it has raised the price of ignoring it, because the fundamentals in question now govern most of the work a junior lawyer will touch. A law school that graduates students without them has outsourced its core curriculum to a first-week seminar at someone else’s firm.
This post draws on Bloomberg Law’s reporting on summer-associate AI training and Law.com’s on AI skills in 2026 summer programs and AI literacy in Big Law recruiting; the ABA Task Force on Law and Artificial Intelligence’s Year 2 Report (December 2025), including the AI and Legal Education survey; UC Law SF’s announcement of its required AI-enabled lawyering lab; Thomson Reuters’ 2026 Law Student Pulse Survey; Law360 Pulse’s reporting on Ropes & Gray’s summer associate AI program, The American Lawyer’s feature on the TrAIlblazer Cup, and Reuters’ reporting on the TrAIlblazers program; the associate-hiring data reported by the ABA Journal; and the duty of technological competence under Model Rules of Professional Conduct r. 1.1 cmt. 8 and A.B.A. Formal Opinion 512 (2024). It builds on earlier posts on the verification standard, three schools’ AI policies, Chicago’s 1L device ban, how lawyers should prompt, the delegation framework, AI prompts as work product, the disclosure patchwork, context-window degradation, and Q1 2026 citation sanctions.