What We Lose When Lawyers Let LLMs Write

One of the first things a bleary-eyed first-year law student has to hear ad nauseam in their first week of law school is that, over the next three years, they will be taught to “think” like a lawyer. What does that mean exactly? And how is that likely to change in the face of rapid adoption and spread of Large Language Models (LLMs) across law schools,¹⁾ law firms,²⁾ and in-house legal departments?³⁾ What do we lose when lawyers let LLMs do their writing for them?

These questions arrive at a fraught time for the legal profession. A recent Thomson Reuters Institute survey found that judges already report declining litigation and communication skills among attorneys in their first five years of practice, with nearly 60% saying client advocacy is being harmed.⁴⁾ Of course, answering these questions depends less on what LLMs can do in the abstract and more on what a given lawyer chooses to do with them, and at what stage of their career. Consequently, the implications of an LLM writing the first draft of the first essay that our harried first-year law student must write in law school differ radically from when, say, it does a final read-through of a share purchase agreement for a third-year law firm associate.⁵⁾ Similarly, delegation will not affect everyone uniformly: a law student incubating their legal judgment through iterative exercises in law school will be affected differently from a more advanced legal professional whose judgment is already settled.

In this symposium post, we unpack how “thinking like a lawyer” is changing in the age of LLMs. At the outset, it’s important to clarify that we understand the term to mean a bundle of specific cognitive acts, comprising the skills of reasoning, argumentation, sequencing, and revision that legal training is built around,⁶⁾ and it is likely that they’re going to change in unpredictable ways once LLMs can do each of these on demand. Of course, some of this lawyerly thinking occurs orally, whether that’s in oral argument in a courtroom, or when an instructor engages in Socratic dialogue with a law student. But we focus here on writing because that is where legal reasoning is captured in a form upon which others rely, and where LLMs are most likely to take over.

We make two arguments. First, we argue that LLM-assisted writing is likely to negatively affect creative lawyering. The common law’s adversarial system is contingent upon a triadic division of cognitive labour: we see modes of what the creativity literature calls ‘divergent thinking’ from advocates putting forth opposing readings of the same factual record, resolved by ‘convergent thinking’ by a third actor — usually a court — that decides which is the winning proposition between the two. When advocates on both sides draft with the help of an LLM trained on the same corpus of published legal writing, that divergence can narrow even while the disagreement looks intact on paper.

Second, we argue that legal writing often constitutes legal reasoning. Legal reasoning does not fully exist until it ‘survives’ the act of documentation, of being written down. As a result, delegating the writing risks delegating away a crucial intermediate step which legal professionals use to ask if the reasoning holds up in the first place.

What happens when we’re all drinking from the same well?

The first aspect of lawyering we address is how offloading cognitive tasks to LLMs affects creative lawyering, or the capacity to think and reason outside the box. Cognitive offloading is the use of an external aid that reduces the mental demand a task places on us.⁷⁾ Think shopping lists or calendar invites.

Creativity research distinguishes two modes of thought: convergent thinking narrows toward a single best answer that derives from a fixed set of premises, which, for example, multiple-choice tests reward. On the other hand, divergent thinking generates the widest possible range of dissimilar responses to the same prompt. Much of the early work on human intelligence treated the latter as central to creativity itself — likely because it was believed that a mind capable of only one answer per question cannot be said to be inventing anything.⁸⁾

In a foundational work in the genre, Carrie Menkel-Meadow argued that good legal creativity requires ‘separating inventing from evaluating’, that is, the ability to generate options before judging them, while saying that much of legal education is aimed at training the critical, convergent half of these two rather than the generative, divergent half.⁹⁾ This leaves law students with a honed capacity to narrow an argument down but impoverishes their capacity to widen the search space in the first place.

The adversarial legal system we inhabit manufactures divergent thinking through design: two advocates work with a (usually) agreed-upon factual record and have access to a similar body of doctrine. Yet, they’re (usually) required to produce maximally different, yet maximally motivated readings of both, before a court performs the convergent operation of choosing between them. The jurist Karl Llewellyn treated this division of labor as one of the ‘steadying factors’ that made common law decision-making predictable.¹⁰⁾

Recent creativity research may help here. A study of LLM-assisted writing found that even when individual stories became more creative and better written, the assemblage of LLM-assisted stories became measurably more similar to one another than stories written by humans alone.¹¹⁾ The gains for individual stories came at the cost of collective novelty. Research on algorithmic monoculture offers an explanation: systems built on shared components tend to produce homogeneous outcomes.¹²⁾ When we apply this finding to two opposing counsels, it shows how our adversarial system’s inbuilt mechanisms to produce divergent thinking are imperiled. What follows is that each side of a legal dispute writes a locally better brief while the ideational distance between the two briefs — the meat on which a convergent judicial decision is supposed to operate — becomes smaller.

Recent work from Finke and Casper finds that models from major LLM developers typically discuss their creators’ controversies in a differentially positive manner.¹³⁾ Let’s imagine a situation where the legal question under consideration implicates the LLM developer’s own exposure, such as fair use considerations for AI training. In such cases, both sides’ briefs risk reading as more defendant friendly.

Choi and Schwarcz’s finding that LLM assistance narrows the spread of legal-writing quality toward a shared mean may be analogous indirect evidence of such a mechanism.¹⁴⁾ The productive tension between opposing sides of a legal argument may become narrower as we see wider adoption of LLM tools, which might play out in unanticipated ways in how the law mediates between opposing parties on things that we deeply disagree about in a plural human society.

Is Cognitive Offloading Domain-specific?

The second question we tackle is whether there is something unique about the relationship between legal reasoning and legal writing, and how LLM-assisted writing affects it. The empirical literature on LLMs and cognitive offloading asks whether such reliance erodes the capacity that it replaces. A 2025 study of essay writing with GPT-4o found that 83.3% of LLM-assisted writers could not remember a correct quotation from their own essay when compared to 11.1% of unassisted writers.¹⁵⁾ Indubitably, a failure to recall one’s own sentences isn’t the same as a failure to grasp their content, so this finding should be treated with caution. Yet, our interest lies in the premise these studies share: that writing involves a separable cognitive process, and that the question is whether outsourcing degrades a process that would have happened anyway. While writing shapes how thinking occurs in virtually every field, the law is a domain where there is a specific difference: while, say, a news story or a lab report describes something that exists apart from it, much of legal text is itself the legal act. Most forms of legal writing, as a result, do not necessarily operate the same way.

Take judicial opinions, for example. On one view, opinions function as rules, which must exist as a text before binding anyone: there isn’t a pre-textual holding waiting to be reported on.¹⁶⁾ Most participants in the legal system work this way. This is not to say that legal reasoning only occurs through the written word: oral arguments in courtrooms and snappy back-and-forth between a law professor and their student would readily disprove that notion. But when legal reasoning binds a set of people, it usually takes a written form.

Before a judge begins to write a decision, they may know how they want to rule but might not have the exact reasoning behind it. A legal argument doesn’t act as a holding until it’s written very clearly to guide the next case. Legal editing therefore carries a lot of value as it points out flaws in reasoning. An argument can sound convincing in theory, but when expressed in words, it can fall apart very quickly. In his new work on AI and judicial opinions, Richard Re makes a similar point, noting how an opinion merges “directive and justification” into one document.¹⁷⁾ As LLMs make writing judicial opinions easy, judges may find it hard to discover flaws in their reasoning through the writing process. This is because fewer drafts will appear unworkable or need to be abandoned. Prior work on cognitive offloading has focused on how overreliance on tools for writing can impact our ability to think independently over time. Legal drafting challenges a key assumption of that work: that reasoning occurs before writing and apart from it. In legal writing, they are intertwined.

One objection is apparent: haven’t judges long delegated their writing to law clerks?¹⁸⁾ Yes, but unlike LLMs, clerks serve as an additional check against opinions rooted in tenuous legal reasoning. In contrast, LLMs will likely return elegant prose when prompted, making judges less likely to edit it substantially and more likely to be anchored to it.¹⁹⁾ Prudential guidelines like the Judicial Writing Manual from the Federal Judicial Center insist that while clerks may assist judges in some of their tasks, “the opinion must always be the judge’s work”.²⁰⁾ Of course, precise empirical work on authorial attribution remains notoriously difficult in this domain.

Our argument tracks an intuitive curricular distinction: 1L legal writing courses are often split between predictive and persuasive writing.²¹⁾ Predictive legal writing (for example, where a first-year law student is asked to write a memo) tells you (or your client) what a court will likely do, in a seemingly neutral voice that weighs both sides, while eschewing the pull toward the answer the writer wants to be true. Persuasive writing is a zero-sum game. It takes the form of a brief that tells a court what it should do, marshalling only the facts and authority that support your client’s position. The object of persuasive writing is to win, and most faculty will admit that even its neutral-sounding Statement of Facts is an advocacy tool. While these two share some scaffolding, they pursue starkly different purposes, which explains why law schools teach them a semester apart. When LLMs draft for us, we forgo a step that can surface a shaky legal proposition, doomed to fail, before a client hears it from opposing counsel.

Moreover, we believe LLMs are likely to fare worse under delegation for two structural reasons. First, they’re trained mostly on published, persuasive-register text, while predictive memos are privileged and essentially absent from training data. A model’s sense of ‘neutral’ therefore comes largely from the performed neutrality of advocacy documents, which looks the same on the page as the real discipline of predictive writing. Second, alignment training pushes models toward hedging, and hedging softens unfavorable predictions more than favorable ones. Notably, this is unlikely to be symmetrical: we predict that briefs drafted by models would be more favorable to the client than attorney-drafted ones, holding authority constant. We believe this isn’t because the model learned advocacy’s performed neutrality, but because hedging erodes unfavorable predictions more than favorable ones.

Conclusion

So, what do we lose when lawyers let LLMs do their writing for them? Surely, we can’t peg it to the production of competent, grammatical prose, nor to an uncanny ability to whip up razor-sharp legal authority, both of which can be done well by an LLM, with newer models doing even better. We hope to have shown that lawyers use writing as a check on their own reasoning. Similarly, when lawyers argue on opposing sides, the productive tension creates the glue of genuine divergence that adversarial adjudication needs to bind itself together. Additionally, the domain-specificity of cognitive offloading has a surprising implication: law may be one of the domains where writing was not a snapshot of thought to begin with, which is why the stakes of outsourcing it look different here than elsewhere.