The impact of AI on research has already led to much reflection in general¹⁾ and, more recently, on legal scholarship in particular,²⁾ including first attempts to measure how far the practice has spread,³⁾ and to examine what happens when generative systems interpose themselves between readers and law review articles.⁴⁾ On a similar note, we have recently argued that the way researchers build their scholarship is itself under threat, and that the response required is institutional.⁵⁾

In this blogpost, we extend that argument to legal research, where it acquires a further consequence. Legal scholarship does not merely report what law is (or should be) but has an impact on the practices that constitute it. It enters legal practice through courts that draw on doctrine, advocates who argue from it, and the education of jurists who will later legislate and decide. This means that AI-assisted legal scholarship can become part of the materials from which later officials learn what the law is taken to be. The problem is that such material may not be backed by any actual recognition or practice.

First scholarship, and the second

The practices through which scholarly judgment is formed are constitutive of research. When a researcher delegates the central acts of inquiry, the output may appear as good or better, but the person who would have been formed by producing it is not (Novelli and Floridi 2026). Law is a demanding case, because doctrinal scholarship has been argued to be at once knowledge of law and a source of it.⁶⁾

We have called the pre-critical practice of scholarship, lived before AI and without suspicion of its own foundations, first scholarship. Second scholarship is the deliberate re-appropriation of that craft by someone who has used AI systems and learned what they can and cannot do.⁷⁾

On this account, four features of research resist (full) delegation and therefore must be protected in second scholarship: tacit knowledge, personal commitment, socialisation, and deep reading. Although resistance to automation is not, by itself, a ground for protection, these four features warrant it insofar as they are the conditions under which judgment is formed and claims are validated. Keeping a human in the loop as a supervisor does not solve the problem because supervision draws on a narrower competence than first-order production. But most importantly, losing these formative conditions compromises the very ability to judge the accuracy of an AI-generated output.

Tacit Knowledge

In law, the tacit dimension has at least one specific object, which is the difference between what the sources say and what is actually in force. A precedent may be formally good and practically dead. A provision may sit on the books and never be invoked. A commentary position may appear in every footnote while carrying little conviction among specialists. Llewellyn captured this discrepancy most directly in his distinction between paper rules and working (or real) rules.⁸⁾ His later notion of situation-sense describes a related practical competence: the trained capacity to apprehend the “type-facts in their context” and their pressure toward a “satisfying working result.”⁹⁾ In Polanyi’s account of tacit knowledge, that competence would operate by enabling the experienced lawyer to integrate accumulated “subsidiary” clues — for example, the treatment of cases — into the “focal” judgment that doctrine and practice have come apart.¹⁰⁾

Legal tacit knowledge is hard for an LLM to capture. Doctrinal desuetude, for instance, may not be explicitly recorded, since published materials can continue to cite a line of authority after practitioners have ceased to treat it as important. The professional judgment that an authority has lost practical force may instead be transmitted through hearings, chambers, and seminars, by people who sort the live question from the settled one without saying that this is what they are doing.

For some authors, there is a jurisprudential reason to resist reducing this knowledge to sources alone. On Hart’s account, the ultimate criteria of legal validity rest on a convergent practice among officials who adopt the internal point of view towards those criteria.¹¹⁾ What counts as law at a given moment is a fact about a practice, not only about a set of texts, and often no quantity of text can fully disclose it.

Present models work primarily from externalised data and representations of legal practice, and they do not occupy the professional relationships in which some clues acquire their significance. Supplying more text does not help if the missing element is the practical weight attached to a legally relevant proposition in a particular institutional setting.

Personal Commitment

Genuine inquiry requires personal commitment, an affirmation in which the person is at risk, and it is that exposure which makes a claim answerable.¹²⁾ Law is not unique in this, but two of its features sharpen the issue. First, legal interpretations are enforced. Cover opens his essay with the line that “legal interpretation takes place in a field of pain and death”,¹³⁾ and that is why answerability here has a hard edge: when a reading is adopted, someone is convicted, deported, taxed, or refused. Admittedly, Cover’s subject is the judge rather than the commentator, but legal scholarship still influences coercive decisions indirectly by shaping the arguments that enter the institutions in which such decisions are made.

Second, legal systems are not there just to issue norms in a void but to maintain institutions that sponsor arguments about them.¹⁴⁾ An institution that sponsors an argument presupposes parties who can be pressed and who have something at stake in the answer.¹⁵⁾ Hence, the absence of such a bearer is the second ground of non-delegability.

Socialisation

Partly connected to the tacit knowledge argument, a third condition is that knowledge, in law as elsewhere, emerges from structured community interactions. Legal interpretation, for instance, is often disciplined by professional norms — Fiss called them disciplining rules¹⁶⁾ — whose authority derives from an interpretive community that recognises them, while that recognition defines who belongs. The interpretive community in which scholars are formed overlaps with, and continually replenishes, the community of officials whose convergent practice grounds the criteria of validity, so the competences reproduced in the former circulate into the practices by which law is identified and applied. Competent participation in practices governed by norms the community deems relevant is therefore essential. Law is unusually well equipped for this, with formal socialisation machinery that most disciplines lack, including traineeships, clerkships, moots, and the bar.

Can an LLM be a member of such a community? Simulated social environments (for example, through multi-agent systems) remain a poor substitute for the thing simulated.¹⁷⁾ But even assuming greater accuracy, membership in an interpretive community is constituted by accountability, and a model trained on (internet) text inherits the outputs of socialised humans without undergoing the formation that produced them.¹⁸⁾

Deep Reading

Deep reading takes a distinctive form in law. What Wolf calls the deep-reading circuit is the capacity, built by use and weakened by disuse, to follow an argument across its structure and to test local claims against the whole.¹⁹⁾ Legal texts are a strong case, as one classic account of precedent shows. On Goodhart’s account, the ratio decidendi of a case, the part of a judgment that binds later courts (as distinct from obiter dicta), is fixed by the facts the judge treated as material together with the decision reached on them, and not by the judge’s own statement of the principle, which may be mistaken without unsettling the precedent.²⁰⁾ The holding is nowhere on the page as a quotable proposition, but the reader reconstructs it from the relation between its elements.

The erosion of deep reading, facilitated by extensive or uncritical reliance on LLMs, leads to two risks. The first, which we have called hermeneutic drift,²¹⁾ is the progressive displacement of a text’s meaning across the user’s iterative prompting of LLMs, which are notoriously deferential. The second extends Kennedy’s phenomenology of a judge assigned a case that seems to set the law against how he wants it to come out, who experiences the materials as a field that resists him.²²⁾ A scholar who uses LLMs uncritically meets no such field, since each question carries the user’s frame and sets the terms of the next. Over successive iterations, this feedback loop weakens the resistance ordinarily supplied by sustained engagement with an established body of legal materials.

The Bigger Picture

A consequence larger than the loss of these scholarly conditions becomes evident. Law is institutional in character, which means that its categories persist through collective practices of recognition and application and produce concrete consequences by carrying deontic powers: obligations, permissions, and competences.²³⁾

Even if scholars are not officials, and nothing a commentator writes makes anything valid law, scholarship supplies materials to the practice that determines validity, through (at least) three channels of unequal formality: (a) directly, where courts treat doctrinal writing as persuasive or, in some systems (far less in common-law systems), quasi-authoritative material; (b) indirectly, through the arguments counsel puts before a court; and (c) through the training of the jurists who will later decide, legislate, and advise. Legal scholarship, therefore, does not merely describe its object but supplies the categories in which the object is stated and, in degrees that vary by system, part of the authority with which it is stated.

It follows that AI-assisted legal scholarship enters this same process, becoming part of the material from which later practitioners and officials learn what the law is taken to be. However, the resulting feedback risk remains contingent and mediated, given that not every change in scholarship changes what constitutes law. To be more analytically precise, we should distinguish at least four levels: (1) perceived scholarly consensus (i.e., what jurists take the state of opinion to be), (2) doctrinal interpretation, (3) patterns of official application, and (4) practices fixing the ultimate criteria of legal validity. A corpus-level shift may alter the first without touching the others, but where it influences counsel, judges, or other officials, it may more or less directly contribute to changes in operative doctrine or official practice and, occasionally, to the practices relevant to validity itself. The possibility is therefore sufficiently plausible to warrant attention: at the margin, scholarship produced without passing through the same formative conditions could lead representation in a textual corpus to function as a proxy for recognition by a community. The proxy, however, would be fallacious as the representation need not be backed by any actual recognition, not least where AI systems recursively summarise prior AI-mediated scholarship, so that a position can gain weight in the corpus without anyone having read it deeply, answered for it, or defended it before a community able to reject it.²⁴⁾

In law, second scholarship is the maintenance of a class of people who hold reasoned views about the law that are not summaries of what has been said about it, and who are therefore positioned to notice when the practice moves. A discipline that outsources its picture of the practice loses the capacity to see the practice change.