Legal AI, Education, and the Risk of Cognitive Conformity
Although justice cannot be reduced to a moral value, it inevitably raises moral as well as normative questions. Over the recent years, one of the most persistent debates in legal theory and practice has concerned the entry of AI into the juridical system. The central question is: can justice become algorithmic?
Judgement, especially in law, does not consist simply in applying rules to a given context. As Dworkin’s account of law as interpretation stresses, it also involves the interpretation of facts, the evaluation of circumstances, and the capacity to reason case by case. The long-standing tension between rules, which secure equity and predictability, and situated judgement, which handles the case the rule does not quite fit. Lawyers learn to manage that tension through legal education, which is why the way law is taught shapes the way it is practised. Yet, the development of AI in the legal sector has largely taken place through assistive tools designed to make legal work faster and more efficient. Recently, AI has spread rapidly across the legal field because of the AI’s systems growing ability to read, summarise, compare, and draft complex documents.
AI can already help lawyers move faster across legal research, document review, due diligence, contract analysis, and drafting. Tools such as Claude Legal support these workflows, without removing the professional responsibility of the lawyer or the judge. Choosing such tools demands attention to accuracy, transparency, security, jurisdictional coverage, and confidentiality. As workflows become more agentic, the lawyer will increasingly supervise the whole process rather than merely correct the final output. This is where governance becomes essential. But governance is only one side of the problem. The other concerns education.
A recent paper entitled Law Professors Prefer AI Over Peer Answers, begins from a practical question: can AI be useful in the education of law students? Its findings open a deeper theoretical issue. Legal education supported by AI may produce a functional convergence between the future legal professional and the legal AI agent, with concrete disadvantages for the former.
Law students trained by AI: A convenient solution?
The authors of the study start from an important methodological problem. Many evaluations of AI work well when there is a clearly correct output. Law, however, often requires judgement, argumentation, and the management of ambiguity. For this reason, the study uses first-year contract law teaching in US law schools as its test case.
Sixteen law professors formulated representative questions, similar to those that students might ask in class or during office hours. The questions concerned cases, doctrine, legal hypotheticals, and policy issues. Both the professors (with access to the course casebook) and two AI products based on the same model, Gemini 2.5 Pro and NotebookLM, answered those questions.
The answers were then compared anonymously. In blind evaluations, contract law professors were asked which answer they would prefer to give to a student who had raised that question. They did not know whether the response had been written by another professor or generated by a large language model (LLM). The evaluators were asked to consider clarity, doctrinal accuracy, and pedagogical usefulness, while retaining space for their overall professional judgement.
The main result is clear: AI-generated answers were preferred over human answers in roughly three-quarters of the comparisons.
According to the authors, the models did not merely produce more fluent or better organised prose. Their advantage appeared to involve the substantive quality of the response as well. The AI answers seemed more accessible to students, less pedagogically misleading, and preferable not only in cases of doctrinal recall but also in exercises requiring legal reasoning. They also appeared to capture a recognisable professional standard.
This finding should be interpreted carefully. The AI systems performed better in defined and circumscribed tasks: answering specific student-facing questions in a controlled evaluative setting. The test case matters, too: contracts is a relatively structured doctrinal domain. It is far from obvious that the result would hold in morally loaded, higher-stakes fields such as criminal law, where judgement carries a different weight. The result is nevertheless significant. It shifts the debate from pure accuracy, to the capacity of AI to approximate expert legal judgement.
The risk of functional convergence
The study measures professors’ preferences, not students’ actual learning outcomes. Yet the paper is interesting because law becomes a suggestive test case for other judgement-rich domains in which expertise depends on decision, interpretation, and argumentative responsibility.
The authors also note that newer models seem to widen the gap between AI-generated answers and human ones. However, if legal education is meant to form judgement, then an AI tutor may alter the way in which judgement itself is learned.
The paper acknowledges that the central task of legal education is the cultivation of judgement: the ability to weigh competing reasons, confront ambiguity, and reach defensible conclusions. What can be deduced from the paper is that a student who learns primarily through polished, professionally acceptable AI answers, may internalise the final format of legal reasoning more readily than the living process that produces it. This does not make the outcome inevitable. In current research on AI in education, the erosion of judgement is read less as a property of the technology, than as a symptom of how it is integrated. Reliance on AI tends to correlate with weaker critical thinking largely through cognitive offloading. Yet guided, reflective use, like surfacing hard cases, making automation bias explicit, testing the reasoning process rather than the polished output, can protect and even strengthen higher-order thinking. The risk, then, is that a student learns only its product and never rebuilds the process that generates it.
That risk is best described as AI becoming an implicit operational model. The student may be trained to reason like an LLM: to recognise patterns, anticipate the most defensible answer, reproduce the average professional standard, and minimise visible errors. As opposed to reasoning like a jurist in training: sitting with doubt, researching first-hand, building an argument from competing reasons, and being wrong in ways that teach. The concern is not the quality of the answer but which capacity is exercised. Reproducing the recognisable form of good reasoning, or developing the judgement that produces it. Exposure to doubt, autonomous research, and the slow construction of an argument, may be reduced.
Ars oratoria vs computational rhetoric
There is also a second risk: the normalisation of the standard. The paper interprets the convergence of professors’ preferences as evidence that LLMs can capture a shared professional standard. If that standard becomes the main educational environment, students may be guided towards what is most recognisable as a “good answer” in an optimised sense. Rather than towards a deeper and more situated form of legal appropriateness. Faced with a borderline case, a student trained on the modal answer may reach for the most-cited rule and the standard argument. When the situation in fact calls for an unusual analogy or a fine distinction that no average-of-the-field output would suggest. At that point, cognitive conformity could narrow the difference between a future legal professional and a legal AI tutor.
This risk is particularly relevant in the case of LLMs used in law. In the legal field, language is part of legal action. Legal language is performative rather than merely informational, in the sense introduced by J. L. Austin — words that do things rather than only describe them. To argue, persuade, order facts, make a claim plausible, modulate tone, and anticipate objections. All of this belongs to ars oratoria. The classical art of oratory, and to the rhetorical dimension the tradition has long placed at the centre of legal practice. Because these systems set the standard for what a “good” legal answer looks like, they may quietly reshape that persuasive and performative craft itself. In teaching students to produce the form of argument the model rewards rather than the argument the situation demands.
The consequence would be a form of functional convergence. The future lawyer learns to produce answers that increasingly resemble those of a well-trained LLM. Ergo, the levelling effect would concern the performance of legal reasoning itself. If an LLM produces an answer that professors prefer because it is aligned with disciplinary standards, the student may treat that form as the model to imitate, and law would then be learned as an optimised output. For this reason, we might speak of a performative levelling of legal judgement.
But there is a difference between formation and conformity.
Legal education should form a subject capable of judgement and exposition. Cognitive conformity would produce a subject capable of replicating formally excellent answers, while being less trained in the critical genesis of those answers. Here, the future lawyer is progressively educated according to criteria that make human legal performance more similar to AI performance.
Nevertheless, AI models can be used to surface disagreement rather than consensus. To expose students to hard cases where the modal answer fails, to make automation bias an explicit object of study, and to assess the reasoning process, rather than only the finished answer. Used that way, an AI tutor trains judgement instead of replacing it. By keeping alive the slower, more uncertain work that produces judgement in the first place: the work that legal education aims to cultivate.
