Over the past year, law firms have bought generative-AI platforms and rebuilt their research and drafting workflows around them. The first-year associate who starts this summer may receive a license to the firm's AI platform in their first weeks and be expected to use it, and the work that used to train them is now the tool’s: the document review, the first draft of an ancillary agreement, the disclosure-schedule check. That work was tedious and billed at a junior rate, and it was also where a new lawyer learned what a deal term looks like when it is market and what a diligence record looks like when it is thin. The tool absorbs the work, and the training that rode along with it.
The question is usually posed as whether to let new lawyers use AI, and on those terms the firm has already answered yes. The better question is what a new lawyer owes for what the tool produces, and there are three answers available. A firm can prohibit junior use of AI, which is unrealistic and, where the prohibition can be evaded, counterproductive. It can hand the tool over and let associates find their own way, which lets them skip the work that builds the judgment the tool’s output requires. Or it can permit AI from the start and require that every substantive output be verified against reliable authority, corroborated rigorously against more than one source. Verification keeps the tool while preserving the part of the work that teaches, and it is the only one of the three that trains an associate for the practice they are entering. Law schools face their own version of the choice, which I have taken up separately; the firm is where the competence either exists in a new lawyer or does not.
Why prohibition fails
Begin with the firm that tries to keep its juniors away from AI until they are ready. The position is hard to hold inside a firm that has already committed to the tool everywhere else. An associate told not to use the tool their practice group runs on is being trained for a version of the job that no longer exists, and they know it, which is why an internal ban tends to drive use underground rather than stop it. The research services compound the problem: the platforms lawyers already rely on to find authority now build generative AI into how they surface cases and statutes, so even a firm that bars the standalone drafting tools cannot wall its juniors off from AI-assisted research without barring research as it is now done. The labor market pulls in the same direction: lateral hiring in the AI specialty rose 68 percent across the Am Law 200, and hiring of associates with AI experience more than doubled year over year, as firms compete to recruit the fluency they want. Surveys of legal professionals, in turn, find that access to these tools shapes whether lawyers stay, so a firm that positions itself as the place where juniors are kept from them recruits and retains against the current.
The pressure the firm itself creates all but guarantees that use: a junior lawyer with a filing due in the morning, two memos already late, and a partner waiting on a turn by the end of the day will use whatever closes the distance between the hours they have and the work they owe, and a tool that produces a usable first draft in minutes is the most direct relief at hand. A prohibition asks them to absorb the overwork and miss the deadline for a benefit they have no reason to credit, and a rule that collides with a critical deadline loses to the deadline. The same prohibition undercuts the competence it means to protect, because that competence now includes the tool. The duty of technological competence under Model Rule 1.1, comment 8, requires lawyers to understand the benefits and risks of relevant technology, and ABA Formal Opinion 512 (2024) applies that duty directly to generative AI. A lawyer kept away from AI through their formative years arrives in practice unprepared for an obligation the rules already impose on them. The citation sanctions that filled the dockets over the past year fell on lawyers who could not tell a real authority from a fabricated one, which is the judgment a prohibition postpones rather than builds.
Why turning new lawyers loose fails
The opposite approach, handing the associate the tool and letting them find their own way, fails for a reason easy to miss when the work product looks good. A junior who drafts their first memos by prompting a model, accepting what reads well, and moving on never does the work that turns information into judgment. The struggle they skip is the part that teaches: building the search that surfaces the controlling authority, reading the cases closely enough to see which holding governs, noticing the adverse decision the model left out. Cognitive psychologists call these desirable difficulties, the effortful retrieval and problem-solving that feel unproductive in the moment and produce durable learning anyway. A tool that returns a competent draft removes the difficulty, and the learning leaves with it.
The profession has begun to register the same problem from the inside. The apprenticeship that trained associates for a century ran on exactly the work AI now does, and as that flow of formative work disappears into the tool, the proxy for training disappears with it. Bloomberg Law and the Thomson Reuters Institute have both described a widening mentorship gap and a coming skills shortage; as Bloomberg Law put it, legal mentorship was already fragile and AI merely made the fracture visible. The associate who never struggles toward an answer does not build the capacity to recognize a wrong one, and the partner who assumed that struggle was happening somewhere may find that it was not.
That capacity is the one the firm most needs as the tools spread, and turning juniors loose is the least likely way to produce it. An associate can turn in clean, well-organized work for years while quietly losing, or never acquiring, the ability to evaluate what the model gave them. The quality of the immediate output, the thing an unsupervised AI workflow optimizes, says nothing about whether the lawyer producing it is becoming someone who can supervise it. I have made a version of this argument about AI tutors that produce answers experts prefer to a human’s without producing better learning; the distance between a good answer and a trained lawyer opens just as wide at the firm.
What verification requires
Verification occupies the space between the two failures. The associate can use the tool from the start, but they owe something for every substantive output it returns. They must corroborate its outputs against a reliable source they consulted independently, reading the case the model cited rather than the model’s summary of it, finding the statute in the code rather than in the answer, confirming the proposition in a secondary source that lawyers in their field rely on. The output is a lead—a starting point—not a final product.
Rigorous verification asks for more than a single confirmation. Corroborating an output against two or three independent sources is how a lawyer catches errors and solidifies their substantive knowledge. An experienced lawyer relies on expertise and intuition that have developed through exposure to numerous sources of information; building triangulation into a junior’s training from the first assignment forces development of the same sort of expertise where the use of AI alone could bypass it.
The requirement reaches every kind of AI output. When AI drafts, the associate confirms that the authorities the draft relies on say what it claims they say. When the model analyzes, they test its reasoning against another source. Even when an associate uses tools with adequate grounding and citation verification features, they must still verify rigorously because grounding may eliminate or reduce fabrication but still risk mischaracterization. Verification is the practice of treating a generated proposition as unproven until an independent source proves it, a discipline that catches more than invented citations and the same competence I have described before as evaluating generated work rather than producing it.
Supervising the verification, not the work
Verification answers the firm’s training problem because it gives the supervising lawyer something to teach and the junior something to learn. When the formative work was drafting or diligence, a partner trained an associate by reading their drafts and marking where the reasoning broke down or checking their diligence memo. As the tool absorbs these menial tasks, the training has to attach to what the associate still does that the tool cannot, and verification is that work. A partner can supervise how a junior confirmed the authorities in a memo as readily as they once supervised the memo itself: which sources the associate consulted independently, how they reconstructed the model’s research without it, where their corroboration ran thin. The apprenticeship follows the work to the verification, where a senior lawyer can watch a junior reason and correct them exactly as before.
This reorientation describes what good supervision of AI required all along. The senior lawyer’s job was never to redo the junior’s work; it was to know the failure modes well enough to catch them. AI’s failure modes look different from associates’. They fail fluently and silently. Teaching a new lawyer to verify is teaching them the failure modes of the tool they will use most, which is the part of supervision that does not change when the tool changes.
What the firm gains from the shift is a training program it can run. Documenting how an associate verified an output, including the provenance of each authority and the independent path by which they could have found it, leaves a record a supervising lawyer can review and correct, in place of the drafting record the tool has taken over. The associate gets the formative work back in a new form, and the partner gets a junior whose verifications they have watched closely enough to trust.
Does verification assume the skill it builds?
The strongest objection to making verification the standard is that it presumes the very skill it is meant to build. A lawyer cannot verify a holding they could not have recognized unaided, and asked to confirm an output they have no independent basis to judge, a true novice will read the model’s summary, find it plausible, and call it verified. On this view verification is less a middle path than a permission slip, letting the new lawyer lean on the tool while performing a check they are not yet equipped to perform.
While a careless version of verification is surely susceptible to this criticism, rigorous verification is not. Reconstructing how they could have found the authority without the model, corroborating it against independent sources, reading the case far enough to see whether its holding governs: these rigorous techniques re-impose the cognitive work the tool removed. Of course, some foundation must exist before there is anything to verify with, and that foundation is built before practice, in law school and before. Law schools must teach (or reinforce) critical thinking, logical reasoning, creative problem solving. A firm cannot assume every entering associate arrives with it, but associates who have these core skills have excelled under conventional training models and will continue to excel under verification-as-training models as well.
For most firms and their clients, the appeal of these tools is measured in hours: a research memo that took a day takes an hour, a week of diligence takes an afternoon. The training that used to ride on those hours does not show up in the same accounting, and a firm can capture the speed while the training quietly lapses. Verification is what keeps it from lapsing.
This post draws on reporting on law firms’ rapid 2026 adoption of generative AI in Fortune; analysis of the associate-training and mentorship gap in Bloomberg Law and the Thomson Reuters Institute; the AI talent-market data reported in the ABA Journal (drawing on SurePoint Technologies’ 2025 State of the Legal Industry report) and retention findings in Clio’s 2026 Legal Trends Report; and the duty of technological competence under Model Rules of Professional Conduct r. 1.1 cmt. 8 and A.B.A. Committee on Ethics & Professional Responsibility, Formal Opinion 512 (2024). It builds on earlier posts on three schools’ approaches to AI, the delegation framework, and answer quality versus learning.