One intuitive boundary for AI-assisted legal work is simple: the lawyer writes the first draft. Before a model gets involved, you confront the blank page, decide what matters, pick the facts, commit to a theory, and build enough of an argument that the model is responding to your thinking instead of supplying it. There is real value in that. I am less sure about what it proves.
The idea now has a judicial endorsement. In Douglas v. Deutsche Bank National Trust Co., decided September 3, the D.C. Court of Appeals struck a brief over four citations that did not exist. Senior Judge Glickman wrote separately to say that fake citations are “the tip of a perilous iceberg,” that asking AI to draft a brief from scratch is a practice supervisors “can (and, I think, should) ban outright,” and that relying on AI to do the thinking behind even a first draft “is an abdication of the lawyer’s duty.” He is right about the property worth protecting, which is lawyer judgment. He proposes a control, which is human-first drafting. The question is whether the control establishes the property.
What going first buys you
The case for the human going first is anchoring. If I ask a model to analyze a problem before I have worked through it, the model proposes the issues, picks the frame, decides which facts carry weight, and supplies distinctions I might not have reached. My task has quietly changed from generating an analysis to reacting to one, and reviewing a proposition is a different act from producing it. A first-draft rule blocks the model from supplying that initial anchor. It forces independent engagement before the model sets the anchor, and it makes it harder to hand over the whole problem and then mistake acceptance for review. Those are good reasons to keep the rule. They are not reasons to let it prove more than it can.
The ethics rule is about judgment, not order of operations
ABA Formal Opinion 512 draws the line somewhere else. It says generative AI “may be used as a springboard or foundation for legal work,” including “generating a draft from which a lawyer produces a legal document,” and that lawyers “may not abdicate their responsibilities by relying solely on a GAI tool to perform tasks that call for the exercise of professional judgment.” The opinion does not make drafting order the test. Its concern is whether the lawyer retains the professional judgment the task requires and remains responsible for the resulting work.
Two lawyers make the point. The first writes a complete draft herself, then gives it to a model. The model proposes a different theory, restructures the argument, shifts which facts matter, and introduces a distinction that becomes the center of the final position. She reviews it, agrees, verifies the law, and files. The second asks a model for a rough draft first, then researches the problem independently, rejects the proposed theory, changes the issues, replaces the authorities, and rebuilds the reasoning to her own conclusion. The first lawyer satisfied the rule and the second did not. Drafting order does not tell you which of them exercised more judgment.
A specimen from my own files
I recently found that I had kept several versions of a piece of my own writing that I cared about getting right. The earliest version was written without AI. Fuller versions followed before the final draft. When I went back to reconstruct the process to illustrate my own experience, I could no longer say which parts of those intermediate versions were entirely mine and which already reflected AI assistance. Had this been client work, it would have been a record-custody miss in my own workflow. It was personal writing, so the miss was inconsequential. Opinion 512 suggests marking AI-produced material stored in client or firm files so later users understand what they are looking at. I had kept no equivalent record of AI involvement. That would at least have made the reconstruction easier. It still would not have told me who originated each idea.
The drafts could answer a narrower question: how far the argument moved. I gave the fuller version and the final version to three model reviewers under the same comparison protocol, without saying why the drafts existed or that AI had been involved. All three saw the same basic thing: the final version was not just a polished version of the earlier draft. The argument itself had changed. Some of my original ideas and facts were still there, but they were being used to make a different point.
That told me something useful, and it also exposed the limit of the exercise. The comparisons could show that the reasoning had moved. They could not tell me whether I had moved it, the model had, or we had arrived there through the back-and-forth between us. Neither could I.
That is easy to lose track of, because AI-assisted drafting moves differently from drafting alone. It can become a rapid brainstorming session: try a framing, see where it goes, reject it, rebuild the argument, and try again. Several substantial changes can happen in the time it once took to work through a single draft. By the end, I still knew where I had started. I could no longer reconstruct where every important judgment along the way had come from.
Specify first, then draft, then test
If judgment is the property, the control that matches it is not human-first specification alone. It is human-first specification followed by independent testing of the consequential choices in the result. Before generation, the lawyer sets down what the problem actually is, which facts control, which positions are open, which constraints cannot move, which decisions the model is not allowed to make, and what would cause the proposed answer to be rejected. An outline, an issue tree, a claim strategy, five bullets. Then the model drafts. Then the lawyer independently tests the choices that matter in the result: whether she determined what authority governs, whether she chose the factual characterization, whether she tested the strongest alternative, whether the final position is hers.
The patent version is the same. A rule that the attorney drafts the claims before the model sees the disclosure proves the attorney produced an initial claim set. It does not prove the eventual scope was her judgment after iteration, or that she was the one who identified the abstraction worth protecting and the prior-art distinction that actually mattered. If those are the judgments to preserve, those are the ones the workflow has to record.
Keep the first-draft rule if it helps you. Ask it to prove only what it can. The first draft tells you where the document began. It does not tell you where the judgment came from.
Lana Akopyan is an IP attorney and independent AI-evaluation researcher. She writes about using AI in legal practice without surrendering judgment, privilege, or the duty of competence at The Agentic Lawyer.
Educational only, not legal advice, and no attorney-client relationship is created. Views are my own. Attorney advertising in some jurisdictions.