PaperlayerView exact revision on arXiv

Source-linked AI summary

The Death of the Legal Author. Authority, Intention, And Law-Creation in the Advent of GenAI

Julieta A. Rabanos, Bojan Spaić

arXiv:2609.31621v1cs.CY

TL;DR

Debates have focused on whether AI should create or interpret law, while the prior question of whether AI outputs can count as law has received less attention. The paper examines that question conceptually, arguing that LLM outputs can count as law in three scenarios when their function is legal.

  • Problem

    Debates largely ask whether AI should create or interpret law, while the prior question of whether AI outputs can qualify as law has received less attention.

    Source: §I · p. 4

  • Method

    The paper conceptually compares intention-dependent accounts of legal authorship with accounts under which an authoritative utterance need not involve intention.

    Source: §I · p. 4, IV.2. Two: Author(ity) without… · p. 21

  • Results

    LLM outputs can count as law in all three scenarios considered when their function is legal.

    Source: §V · p. 31

  • Takeaways & Limitations

    The paper’s conclusion supports the possibility that legal authorship and law-creation need not depend on an intentional AI author.

    Source: IV.2. Two: Author(ity) without… · p. 21, §V · p. 31

    Without legal fine-tuning, generated legal-like texts are overly generic and heavily dependent on the human prompter’s abilities.

    Source: III.2. Creating v. Generating legal… · p. 16

Abstract

from arXiv · show

Generative artificial intelligence in the form of chatbots based on large language models (LLMs) has taken the world of law by storm. Philosophy of law is struggling to catch up with the theoretical significance of the advent of technological development and the way it may modify traditionally established understanding of legal phenomena, such as law-creation and authority. In this sense, for the most part, heated philosophical debates have circled around a normative question: "Should AI create and interpret law?". Much less attention has been given to a different, albeit previous, question: "Can AI create and interpret law?" That is, is AI capable of producing outputs that can be deemed as "law" (at least, law as we know it)? Can AI be a "legal author"? This paper explores this unattended question and endeavours to provide some provisional answers. In the first part, we define legal authority, legal authorship, and legal interpretation and claim that the intention of a determinate authoritative author is often considered the condition of the possibility of creating and interpreting contemporary legal texts. In the second part, we argue that LLM AI, in general, and ChatGPT, in particular, generate legal texts without having any intention. In the third part, we consider the positions of the authors that downplay or even eliminate intention from the discussions about the legal character of prescriptive texts. Finally, we argue that there are good reasons to side with the second group of authors. The ability of agents without intentions, like ChatGPT, to create legal text is an argument in favour of the thesis that law can be created without intention behind the creation and that nonintentional creation can be interpreted to arrive at legal norms.

I. INTRODUCTION

The paper shifts attention from whether AI should create and interpret law to whether its outputs can qualify as law and AI as a legal author. It takes up that conceptual question by examining authority, authorship, intention, and legal text.

I. INTRODUCTION

  • ChatGPT has been used to draft legislation, while AI tools are also used in legal practice and can serve as semi-autonomous drafters of legal texts.Examples include a Porto Alegre legislative act written by ChatGPT and a Massachusetts bill drafted with its help.Source: §I · pp. 2–3
  • The paper asks whether AI can produce outputs that count as law and whether AI can be a legal author, questions less attended than whether AI should create or interpret law.The distinction matters because the paper treats the capability question as prior to the normative debate.Source: §I · p. 4
  • A descriptive route tests language models, but the paper takes a conceptual route grounded in prevailing accounts of legal authorship, authority, and intention.Those accounts commonly treat law-creation as requiring an authoritative legal author and a text in which that author's intention is materialized.Source: §I · p. 4
  • The paper offers provisional conceptual answers, assessing whether law-creation can be understood without intentionality or primary emphasis on the creator.Its roadmap examines authority and legal authorship, LLM functioning and outputs, and whether those outputs can count as law.Source: §I · pp. 4–5

II. LEGAL AUTHOR(ITY): FIRST TAKE

The paper approaches legal authority through law-creation, treating it as a relation among an author, an output, and an addressee. In law, a subject’s authority to produce legal effects depends on the power-conferring norms of a particular legal system.

II. LEGAL AUTHOR(ITY): FIRST TAKE

  • Authority can be analyzed through law-creation as authorship of objects produced by certain means, with an author, an output, and an addressee as its basic structural parts.The addressee receives the output and may help shape its characteristics through interpretation.Source: II.1. The Authority: A brief overview… · pp. 5, 7, 8
  • The paper’s reconstruction of authority includes a subject’s capacity to issue prescriptions intended to guide addressees’ actions.It identifies relational quality, capacity, utterance type, and intention as elements of authority, with norm-dependence added for legal authority.Source: II.1. The Authority: A brief overview… · pp. 6–7
  • Legal authority depends on a power-conferring norm in a particular legal system, which sets the power, its conditions, and the effects of its exercise.The system’s criteria determine which subjects can produce legal texts or norms through specified acts under specified conditions.Source: II.1. The Authority: A brief overview… · p. 7
  • A central open question is whether an author must factually intend an output to count as law, or whether that intention can merely be presupposed in context.A related question is whether addressees are passive recipients or shape the output’s characteristics through interpretation.Source: II.1. The Authority: A brief overview… · p. 8

II.2. The Author: Law as an intentional creation

The section presents a familiar account of law as intentional creation: legal texts are typically produced by authorized authors seeking to regulate behavior, and artefacts are commonly understood as purposefully made. It also notes that competing theories differ over intention’s role, while both camps still treat intention as necessary to artefactual status.

II.2. The Author: Law as an intentional creation

  • Statutes, contracts, and other legal sources are often created by determinate, normatively authorized authors intending to regulate human behavior.The section connects this view to the conception of law as an artefact made intentionally for a purpose.Source: II.2. The Author: Law as an… · p. 8
  • On the classical account, an artefact is made intentionally by an author for a purpose, making intentional creation by a determinate author a condition of artefactual status.For Hilpinen, intentional production is authorship and its product is an artefact.Source: II.2. The Author: Law as an… · p. 8
  • Intentionalist and non-intentionalist theories disagree about whether intention determines artefact function or whether use-history and context do, but both require intention for artefactual status.The difference concerns whether intention is sufficient, not whether it is necessary.Source: II.2. The Author: Law as an… · p. 9
  • Many legal philosophers describe law, apart from exceptions such as customs and some moral norms, as created by officials’ purposive actions.The section cites divergent accounts that converge on this characterization.Source: II.2. The Author: Law as an… · p. 9
  • On the intentionalist account of legislation, a legal authority creates law by expressing an intention that a particular text count as law.The account treats intention to legislate as constitutive of the law-creating act.Source: II.2. The Author: Law as an… · p. 10

II.3. The Reader: Understanding authorial utterances

Many accounts treat intention as central to both making law and interpreting legal texts, though they differ on whether intention is a normative guide or a condition of intelligibility. The section distinguishes several kinds of intention that need not entail one another.

II.3.1. Intentional creation and intentional interpretation

  • Intentionalist accounts treat a creator’s intention as necessary to interpret legal texts, sometimes as both a guide and a condition for identifying meaning.Raz’s account also gives intention normative priority; other intentionalist accounts treat it as constitutive of interpretation.Source: II.3.1. Intentional creation and… · pp. 10–11
  • Fish and Alexander and Prakash argue that understanding purposively produced objects presupposes intention, but this does not itself prescribe a legal interpretive method.Fish distinguishes this theoretical claim about intelligibility from claims about legal interpretation.Source: II.3.1. Intentional creation and… · p. 11, II.3.2. Kinds of intention · p. 11

II.3.2. Kinds of intention

  • A communicative account treats legal authority as a speaker whose intention helps constitute a directive’s meaning and guides its interpretation.The account draws a parallel between legal directives and ordinary conversation against a mutually known context.Source: II.3.2. Kinds of intention · pp. 11–12
  • Legislative intentions may concern a text’s meaning, making it law, or its effects; type, token, and causal intentions likewise distinguish creating a kind of object, a particular text, and desired behaviour.These kinds do not necessarily entail one another: a legal text may be created without an intention to make law or change behaviour.Source: II.3.2. Kinds of intention · pp. 12–13

II.4. Summing Up

The literature review finds that legal creation and interpretation are often understood to depend on intention. Generative AI makes nonintentional production of legal-like texts newly salient to these questions.

II.4. Summing Up

  • The review identifies intention behind law as a commonly presumed condition of both law-creation and law-interpretation.The authors describe these as conclusions from their revisitation of the literature.Source: II.4. Summing Up · p. 13
  • Generative AI changes the debate by making legal-text production without an apparent intention newly imaginable.The paper notes that law-creation and interpretation had not generally been considered activities for hardware or software to perform.Source: II.4. Summing Up · pp. 13–14
  • The review distinguishes type and token intentions concerning law-creation from causal intentions concerning law’s effects in reality.The exact output of law-creation—texts, norms, or both—remains an open question in the discussion.Source: II.4. Summing Up · p. 13

III.1. How did we get here?

The paper traces legal AI from early work on legal reasoning and knowledge representation to machine learning and today’s LLMs. It identifies the transformer architecture as a foundation for current language models.

III.1. How did we get here?

  • Legal AI research moved from case reasoning, knowledge representation, and deontic concepts toward machine learning as a prominent technology in the last decade.The paper also notes a turn toward modelling legal knowledge and argumentation schemes as electronic legal information grew.Source: III.1. How did we get here? · p. 14
  • The 2017 transformer architecture became a foundation for current LLMs by using words and their contexts to process and generate language.The paper names GPT, Claude, LaMA, and Gemini as examples of models based on this development.Source: III.1. How did we get here? · p. 14
  • The paper describes LLM uses in legal research, contract analysis, document drafting, outcome prediction, and chatbot assistance.These uses include summarizing complex legal texts and identifying relevant cases, statutes, or arguments.Source: III.1. How did we get here? · p. 14

III.2. Creating v. Generating legal texts

The paper contrasts legal reasoning and drafting capabilities reported in prior studies with the limits of LLM-generated statutes. Its example and discussion conclude that these models can produce convincing legal-like text, but current outputs are generic and prompt-dependent.

III.2. Creating v. Generating legal texts

  • Prior studies described GPT-4 as performing well on exams requiring text understanding, interpretation, and short-form writing, with one study reporting lawyer-standard performance on complex legal tasks.The paper also surveys work on legal reasoning, statutory interpretation, and bar exams.Source: III.2. Creating v. Generating legal… · p. 15
  • Fine-tuned models can produce legal documents, especially in private law, while commercially available models remain limited in law-drafting capabilities.The paper identifies the tested model as the February 27, 2024 version of ChatGPT based on GPT-4.Source: III.2. Creating v. Generating legal… · p. 15
  • The sample statute prohibits carrying firearms in public places and includes exemptions, penalties, enforcement provisions, severability, and an effective date.The draft uses placeholders such as [Your State/Country] and [number of days].Source: III.2. Creating v. Generating legal… · pp. 15–16
  • ChatGPT can generate legal-like texts that are hard to distinguish from human-created legal texts, but the paper stops short of saying they count as law.The authors say outputs from simple prompts are overly generic and heavily dependent on the human prompter, and that expanding the sample statute does not dramatically change it.Source: III.2. Creating v. Generating legal… · p. 16

IV. LEGAL AUTHOR(ITY): SECOND TAKE

The paper argues that current LLMs lack intentions despite producing outputs comparable to human legal creators, and that intention-based criteria therefore exclude their outputs from law and legal interpretation. This tension leads the authors to ask whether human intention should be ascribed to LLM outputs or whether intention is not necessary for law-creation and interpretation.

IV. LEGAL AUTHOR(ITY): SECOND TAKE

  • LLM outputs can sound reasonable without indicating an internal life; the authors cite evidence that emergent behavior is outward-facing and that no current AI systems are conscious [85].They also note that term frequencies in training data may partly explain reasonable-sounding outputs.Source: §IV · p. 17
  • Current LLMs lack mental states or intentions, although their outputs can resemble those of a human law-creator and display relatively autonomous reasoning.The authors report no technical limitations that would in principle prevent LLMs from performing law-creation or law-interpretation tasks.Source: §IV · p. 18
  • Under the paper’s intention-based criteria, ChatGPT outputs cannot count as law or be interpreted as law because no intention can be identified behind them.This applies the paper’s stated requirements for legal texts and legal interpretation to LLM outputs.Source: §IV · p. 18
  • Real legislative uses of LLMs make the exclusion counterintuitive, so the authors frame two alternatives: ascribe human intention to outputs or reject intention as necessary for law-creation and interpretation.They cite instances of legislators using LLMs to draft bills, including one that passed and became part of a legal system.Source: §IV · p. 18

IV.1. One: His Master’s Intentions88

The authors consider whether human intention can be carried into AI-generated legal texts through three scenarios, from AI-assisted drafting to autonomous generation. They distinguish human intention behind a prompt from intention behind the generated words, and conclude that the autonomous scenario removes both authorial intention and its use in interpretation.

IV.1. One: His Master’s Intentions88

  • In the extended-legal-mind scenario, AI supplies and interprets research data for a human drafter, who remains the legal author and uses AI as a tool.The source of law is informed by AI rather than authored by it.Source: IV.1. One: His Master’s Intentions88 · p. 19
  • Even when a human reviews AI-generated material, some portions may lack token intention and thus fail the intentional-author criterion for counting as law.The human drafter has type intention, but the text may contain both human-drafted and AI-drafted parts.Source: IV.1. One: His Master’s Intentions88 · p. 19
  • In the prompt-driven scenario, a human’s law-creating and result-directed intentions make production intentional, but the generated words lack token intention, preventing interpretation through the legislator’s actual intention.The human supplies the prompt while AI produces most of the legal text.Source: IV.1. One: His Master’s Intentions88 · p. 20
  • The autonomous-generation scenario is highly speculative, but it removes human prompting and leaves no human authorial intention to borrow.The paper notes that human beings might still need to interpret law created by such an author.Source: IV.1. One: His Master’s Intentions88 · pp. 20–21
  • Because autonomous AI lacks mental states and intention, the established view would deny its outputs the status of law and deny that interpreting them counts as legal interpretation.In this scenario, attempts to borrow or recover a real intention behind the text would fail.Source: IV.1. One: His Master’s Intentions88 · p. 21

IV.2. Two: Author(ity) without intentions

The authors examine whether legal authorship, authority, and interpretation require an intentional author, considering unintentional authorship or authority and the role of readers and normative practice. They argue that AI outputs can function as legal artefacts when used to regulate human behaviour, without requiring an intentional author.

IV.2. Two: Author(ity) without intentions

  • The authors propose setting aside the search for a human prompter’s borrowed intention and ask whether law can be created and interpreted without intention.They identify this as an alternative to accounts that treat intention as necessary for legal creation, authority, or interpretation.Source: IV.2. Two: Author(ity) without… · pp. 21–22

IV.2.2. The Unintentional Author

  • Artefacts can function without authorial intention when use attributes and maintains their proper function, allowing AI outputs to count as legal artefacts when used as law.The relevant use is legal regulation of human behaviour; the authors argue that use function, rather than creator design, can determine an artefact’s identity over time.Source: IV.2.2. The Unintentional Author · pp. 24–25

IV.2.3. The Unintentional Authority

  • For Olivecrona, legal norms are “independent imperatives”: formalities and the label “law” let them function without an author’s will or intention.Their content and scope depend on addressees’ evaluations, which may be shared yet remain uncertain and change over time; disagreement about meaning is a feature his descriptive account can accommodate.Source: IV.2.3. The Unintentional Authority · pp. 25–27

IV.2.4. The Intentional Reader

  • Poggi argues that legislation involves a shared intention to follow a lawmaking procedure, but not a Gricean communicative intention to express a particular norm.The shared intention to approve a text suffices to establish its assumed normativity and legality, while attempts to reconstruct a legislative intention expressing a specific norm are rejected.Source: IV.2.4. The Intentional Reader · pp. 27–28

IV.2.5. Dispensing of intention

  • Poggi’s account retains a shared assumption of norm-creating intention because legal participants rely on it to describe and explain their interpretive practice, even if no such intention exists.This distinguishes Poggi from Olivecrona, who appears to explain legal practices without reference to legislative intention or addressees’ ideas about it.Source: IV.2.4. The Intentional Reader · pp. 28–29
  • The common strategy is to shift emphasis from authorial intention to readers’ or communities’ intentions, while treating legality as conferred through the impersonal process of law production.The authors present this combination as a way to address problems that Alexander associates with eliminating intention.Source: IV.2.5. Dispensing of intention · p. 29
  • An output’s legal character does not depend on its author’s intention: it is enough that the output performs the function assigned to it by its subjects.Users may imagine a real or fictional author, but that is optional if the utterance’s function is consistently maintained.Source: IV.2.5. Dispensing of intention · p. 29
  • An AI’s lack of intention need not prevent it from acting as a legal authority, because legal status depends on conformity with a normative system’s production requirements, not on an authority labeling its output law.A possible consequence is that the utterance’s content may seem unacceptably indeterminate.Source: IV.2.5. Dispensing of intention · p. 29
  • Recipients can determine a text’s use, confer authority on its generator, and ascribe normative content, so real or imagined authorial intention is not necessary for determinate legal meaning.Judges already choose among possible textual meanings to assign normative content; intention may inform that choice but is not required.Source: IV.2.5. Dispensing of intention · pp. 29–30

V. SOME CONCLUDING REMARKS

A thought experiment mimics AI-generated law through a tribal rulebook adopted for governance, followed by a text assembled from cats’ typing without authorial intention. The paper concludes that LLM outputs can count as law in three scenarios when their function is legal.

V. SOME CONCLUDING REMARKS

  • The tribe adopts a book of prescriptive sentences for social regulation, then consults it to resolve disputes and guide behaviour.The thought experiment asks readers to treat the book as law through its use in communal life.Source: §V · p. 30
  • Cats produce sensible text for a human editor to collect, but neither the cats nor the editor intends to create a legal rulebook.The editor selects and binds the text, while the cats’ typing lacks authorial intention to regulate tribal life.Source: §V · pp. 30–31
  • The thought experiment proposes three routes to legal character: creation without intention, function independent of a creator, or readers’ shared assumptions about an output.The paper’s question is whether AI outputs can count as law, not whether they should be recognised as law.Source: §V · pp. 30–31
  • LLM outputs can count as law in all three scenarios when their function is legal.The thought experiment suggests legal status need not depend on intention behind a text expressing legal norms.Source: §V · p. 31
Loading 2609.31621v1…