The First Draft Is Where the Lawyer Is Made.

This was not a firm with no AI policy

The record makes the incident more instructive than a familiar story about a lawyer casually pasting chatbot text into a filing. On September 22, Judge Subramanian directed lead counsel for Foundation Media to explain whether AI had been used to prepare a response and whether inaccurate quotations and a citation resulted from unchecked AI use. The underlying case is a copyright dispute brought by a self-represented plaintiff. The AI issue concerned a filing within that litigation, not a ruling on the copyright merits.

Lead counsel Cynthia Arato responded under penalty of perjury. Her declaration states that Shapiro Arato Bach had licensed generative-AI tools designed for legal work, conducted several onboarding sessions, required mandatory training from outside ethics counsel, and made employees sign an acceptable-use policy requiring validation of AI output. Everyone responsible had been told to check the work.

The process still failed. According to the declaration, an unnamed lawyer used AI in preparing the response, intended to verify the citations, checked factual citations, and did not complete the case-citation review. One cited Zhao case did not exist. A Piazza decision did not support the propositions attributed to it. Lead counsel knew the legal principles and edited the draft without independently checking those authorities because she did not know AI had been used. The firm acknowledged the failure, apologized, and said it was adding safeguards.

The declaration also disclosed serious personal circumstances affecting the lawyer who prepared the filing, including a recent parental death and another setback. Those circumstances were offered as context, not excuse. They matter because systems should be designed for real people working through grief, fatigue, deadlines, and divided attention. A control that works only when every professional is at full capacity is not much of a control.

A policy can assign responsibility without creating capacity

Most firms now know the basic governance language. Verify citations. Protect confidential information. Use approved tools. Maintain lawyer supervision. Do not let a model make the legal decision. ABA Formal Opinion 512 frames generative AI through existing duties of competence, confidentiality, communication, candor, supervision, and reasonable fees. None of that is controversial.

But a policy describes the required result. It does not necessarily create the time, habit, or skill needed to reach it. If a workflow generates a polished draft in seconds and places verification at the end, checking becomes a discrete cleanup task competing with the filing deadline. Familiar propositions feel safe. Plausible citations survive because the reviewer recognizes the rule and stops asking whether this particular authority actually says it.

Traditional first-draft work imposed useful friction. A lawyer had to define the question, locate authority, decide what each case was good for, and build a structure before the prose looked finished. That process was inefficient in the narrowest sense. It was also diagnostic. A supervisor could see whether the junior lawyer misunderstood the standard, skipped a factual gap, relied on weak authority, or buried the best argument. When AI supplies the structure and prose first, some of that evidence disappears.

This is why more training about hallucinations will not solve the entire problem. The risk is not only that a model invents a case. It is that a junior lawyer becomes excellent at reviewing completed language without developing the ability to produce and defend the underlying analysis. Verification can catch a bad citation. It cannot reveal a missing argument that neither the model nor the reviewer thought to ask about.

Clients can buy efficiency and still receive less value

The judge's warning also reaches clients. Reuters reported that he cautioned clients to think carefully before requiring AI for brief drafting merely to save money, while noting that there was no indication Foundation Media had made such a demand. The observation deserves attention. Procurement pressure increasingly treats AI use as a discount mechanism: if the tool makes the work faster, the bill should be lower.

That logic is understandable, particularly when firms bill by the hour. Clients should not pay for avoidable duplication or mechanical work that technology can perform reliably. But legal service is not manufactured from interchangeable minutes. Some hours produce the immediate document. Others produce the lawyer who will recognize the decisive issue in the next matter. A client that demands the removal of every training hour may reduce today's invoice while relying on a talent system it is helping to hollow out.

Firms cannot answer that concern by hiding inefficient staffing behind the word mentoring. Clients are entitled to know who is doing the work, how technology is used, what review occurs, and why the proposed team is proportionate to the problem. The commercially honest answer is not that juniors must spend unlimited hours reinventing routine documents. It is that some supervised work has two outputs: the filing and the professional capacity needed to produce sound filings later.

Alternative fees make the tradeoff easier to see. Under a fixed or value-based fee, a firm can use AI to reduce production cost while deciding internally how much of the saved capacity to reinvest in training and review. Under hourly billing, those investments are visible and harder to explain. Either way, the firm should not let the billing model make professional-development policy by accident.

Not every first draft deserves protection

A universal rule requiring every junior lawyer to draft everything without AI would convert a useful warning into ritual. Repetitive chronology sections, standard discovery correspondence, first-pass document summaries, and formatting work may be poor uses of developing judgment. A lawyer can also learn by comparing an independent analysis with a model's answer, interrogating the differences, and revising both. The point is not to preserve drudgery. It is to preserve the cognitive work that drudgery sometimes contained.

The right unit is the skill, not the document. Can the lawyer frame the issue before prompting? Can the lawyer find and read the controlling authority? Can the lawyer explain why one fact changes the rule's application? Can the lawyer build an argument from a blank page when the model's framing is wrong? Can the lawyer identify what is absent, not merely what is false? If those abilities are never tested without assistance, the firm does not know whether it has augmented judgment or replaced practice with dependence.

Firms should designate protected repetitions for developing lawyers. A first-year associate might prepare an issue outline and authority map before using AI. A midlevel could write selected argument sections unaided, then use the tool for counterarguments or compression. A supervising lawyer could review not only the finished draft but the research path, discarded authorities, and strategic choices. As competence grows, the restrictions can narrow. That is apprenticeship with instrumentation, not nostalgia.

The missing control is a training architecture

The Hill episode shows why an acceptable-use policy is only one layer. A firm also needs technical controls that identify AI-assisted content where appropriate, citation checks that do not depend on the drafter remembering a final step, review responsibility assigned by name, and enough time between draft and filing for the check to be real. Those controls protect the client in front of the firm.

A training architecture protects the clients who come later. It should define which foundational skills each lawyer must demonstrate, which tasks may be AI-assisted from the start, which require an independent attempt, how supervisors will give feedback, and how the firm will test for overreliance. Usage data can show whether a tool is popular. It cannot show whether a lawyer is becoming more capable unless the firm also evaluates unaided reasoning and supervised performance.

Judge Subramanian did not impose sanctions and did not order law firms to ban AI from first drafts. His warning is more useful because it identifies a problem discipline rules do not fully solve. The profession has spent years asking whether AI will replace junior lawyers. The more immediate risk is quieter: junior lawyers remain employed, produce more polished work, and receive fewer repetitions that teach them why the work is right. A firm can automate a draft. It cannot automate the experience of becoming responsible for one.

Efficiency removes time from a task. Leadership decides whether it also removes the learning.