Why Legal Translation May Be AI’s Strongest Domain

Legal translation is usually presented as the place where machines will fail. The stakes are high, the terminology is treacherous, and a plausible mistake can cost someone a right, a property or their freedom. I accept all of this. Yet I have come to the opposite conclusion: law may be the domain in which large language models are best placed to deliver good translation.

The argument is simple to state. Law is made, to an unusual degree, of language. Its objects (rights, duties, contracts, offences) exist because they are enacted, written and recognised in language. And the knowledge needed to understand those objects is largely stored in language: in statutes, judgments, definitions and commentaries. A large language model is a system built from, and operating on, exactly this kind of material. In most domains there is a gap between what a language model has access to, which is text, and what the task is ultimately about: bodies, machines, perceptions, the physical world. In law, that gap is narrower than anywhere else I know.

Let me be precise about the kind of claim this is. I am not saying that machines already translate law better than expert lawyer-linguists. That is an empirical question, and today the answer is probably no. My argument is conditional, and it has two premises. The first concerns what legal translation is: the transfer of meaning that is constituted and documented in language. As David Mellinkoff put it in the opening line of The Language of the Law (1963), "The law is a profession of words". (See also Susan Šarčević, New Approach to Legal Translation, which treats legal translation as communication within the mechanism of the law itself.) The second concerns what LLMs can do: model language, and reason across very large bodies of text. (See Piantadosi and Hill on how language models can capture meaning through the relations between concepts, even without direct reference to the world, and LegalBench by Guha et al. for measurable legal reasoning in LLMs.) If both premises hold, the conclusion follows, whatever any particular system can do this year. Anyone who wants to reject the conclusion has to reject one of the premises: show that legal meaning lives somewhere other than in texts, or that language models cannot, in principle, work with texts at this depth. Pointing to today's errors is not enough.

“

In law, that gap is narrower than anywhere else I know.

The rest of this post defends the two premises, and then asks what they do and do not imply.

Law is made of language

Most of what we translate is language about something. A medical report describes a body; a technical manual describes a machine; a novel evokes an experience. The text points beyond itself, and a translator who knows only the text is always at one remove from what the text is about.

Law is different in kind. A legal text does not only describe the world; it changes the normative world. When a parliament enacts a statute, a right comes into existence. When two parties sign a contract, an obligation exists that did not exist a moment before. When a judge pronounces a sentence, a person's legal status changes. J. L. Austin called utterances of this sort performatives: in the right circumstances, saying is doing. ¹

Performativity matters because it moves the object inside language. A medical report describes a body that exists whether or not the report is written. A contract does not describe an obligation that exists elsewhere: the obligation exists because the contract was signed, and its content is whatever the text says it is. So in law, the thing the text is about is not on the far side of the text. The text creates it and fixes its content. That is why the translator of a legal text is not at one remove, and why a system that has only text is not cut off from the object.

John Searle's account of institutional facts gives this a general form. Rights and duties are not physical objects. They exist because we collectively accept rules of the form "X counts as Y in context C", and those rules are formulated, stored and applied in language. Searle went as far as calling language the fundamental institution, the one on which all others depend. If he is right, law is its clearest case: an institution whose objects are, to a large extent, made of words.

A reader of Searle will object at once, and rightly. For Searle, institutional facts depend on collective acceptance, on human beings who treat X as counting as Y. A language model takes no part in that acceptance. Searle is also the author of the Chinese Room, the most famous argument that manipulating symbols is not understanding them. But my claim does not need the model to understand law in Searle's sense, still less to create legal facts. Legal facts are constituted by people and institutions. My point is about where those facts are recorded. Collective acceptance in law does not float freely in people's heads; it is fixed in authoritative texts, precisely so that it can be consulted, cited and contested. The translator's task is not to bring legal facts into being but to carry the record of them faithfully into another language. That is a task performed on language, and it is the record, not the acceptance, that a model can read.

This matters for translation in a precise way. I have written elsewhere about the shadows of reality in language. Words carry traces of human experience without being that experience, and language models learn the traces without the experience. That is a real limitation whenever meaning lives in the world outside the text. In law, however, the shadow is often the thing itself. The authoritative text of a statute is not a description of the law; it is the law, or the closest thing to it we have. A machine reading that text is not looking at the shadow of a legal object. In an important sense, it is looking at the object.

“

In law, however, the shadow is often the thing itself.

I want to be careful here. Law is not only text. A court must establish facts about real people and real events; rules are made and enforced by institutions that have power; legal language is interpreted within social and political life. A machine reading a judgment does not know what happened in the courtroom. Nor is all legal knowledge written down: practitioners carry habits of reading and drafting that no statute records. But the legal meaning of that judgment, what it decides, which rule it applies, what it changes, is to a remarkable degree contained in language, and the legal profession has strong incentives to keep it there, because what is not written cannot easily be relied on.

Legal knowledge is encapsulated in language

The second half of the argument concerns knowledge. Legal meaning is rarely found in an isolated word. It is found in the relations between texts.² A contract refers to legislation; legislation defines terms used by other legislation; a judgment interprets a provision in the light of earlier judgments; a commentary organises the judgments into doctrine. Katia Peruzzo describes this intertextual network as a defining constraint of legal translation: the document in front of us belongs to a wider body of legal texts, even when most of that body will never be translated.

H. L. A. Hart observed that legal concepts have an "open texture": a settled core and a penumbra of doubtful cases.³ What fills the penumbra over time? More language. Decisions, reasons, distinctions, dissenting opinions. Each hard case that is decided leaves a written trace that narrows or extends the meaning of a term. Over centuries, legal systems have built something rare among human practices: a body of knowledge in which a very large share of the relevant distinctions has been argued out explicitly and recorded in writing.

One might reply that open texture is exactly where text runs out: the next hard case has not yet been decided, so no corpus can contain its answer. That is true, but it is a problem for the judge, not for the translator. The translator is not asked to settle the penumbra. The translator is asked to carry it across intact: to render a term so that what is settled stays settled and what is open stays open, with neither more nor less room for interpretation than the source allows. The same holds for ambiguity that drafters have left in a treaty or a negotiated contract on purpose. A good legal translation does not resolve what the parties chose not to resolve. To do this, one needs to know where the boundary between core and penumbra currently lies, and that boundary is documented in the very texts that drew it.

Niklas Luhmann described law as a self-referential system, one that reproduces itself through its own communications. I would not push the analogy too far. Law is not a closed game like chess. But the way law makes facts legally relevant runs through its own documents, categories and procedures. In this respect it resembles software more than literature: to understand a line of code, you inspect the system in which it runs; to understand a legal sentence, you inspect the body of law in which it operates. Gadamer went further still, treating legal hermeneutics as the model for all understanding, because the jurist must always apply a received text to a new situation. The legal translator does something very similar: carry a text into another language, and often into another legal order, while preserving what that text does.

Put the two halves together and the conclusion follows. If the objects of law are constituted in language, and if the knowledge needed to interpret them is encapsulated in language, then a system whose whole competence is the processing of language across vast bodies of text is not an outsider to law. It is working in law's own medium. This is why I think LLMs should be, in principle, better suited to legal translation than to medical, technical or literary translation, where much of the decisive knowledge lies outside the text.

The obvious objection is literature. A poem is made of language too, perhaps more completely than a statute, and nobody claims that poetry is the natural domain of machine translation. The difference lies in where meaning is settled. The meaning of a poem is completed in the reader: in experience, memory, rhythm felt in the body, an effect that no further text can fix. A legal text is designed for the opposite. Its meaning is meant to be settled by other texts, through definitions, cross-references, precedents and authoritative interpretation, so that two readers far apart in time and place can arrive at the same consequence. Literature is made of language and points out of it, into experience. Law is made of language and points back into language. That is the property that matters for a language model.

“

Law is made of language and points back into language.

Translating between legal systems

Legal translation does not stop at understanding the source law. The reader often belongs to another jurisdiction, where a familiar-looking word names a different institution, or where no equivalent institution exists at all. There is no neutral legal metalanguage standing above national systems into which every concept could be translated first. The Court of Justice of the European Union describes exactly this problem: concepts developed in one legal system must be made intelligible in others, and a term in the target language may have no direct counterpart. Its lawyer-linguists use comparative research and multilingual terminology records that capture not only preferred terms but definitions, sources, legal-system origin and warnings about misleading similarities.

Note what this work consists of: comparison of texts, documented in more texts. Here again the knowledge is in language, and this is where I see a strong role for reasoning systems. Rather than asking a model How do I say this word in French?, we can ask it to trace what the source term means in this document and under this jurisdiction. Which definition governs it? Has a court narrowed its meaning? Does the target system contain a comparable concept, only a partial one, or none? Is the translation meant to have legal effect, or only to inform a foreign reader about the source law? These questions can lead to different, equally legitimate choices. The goal is not to make two legal systems look identical. It is to make their differences visible where those differences matter.

A system can then propose a rendering, check it against both legal systems, and explain its choice. It can flag a tempting false equivalent, keep a defined term consistent across a long document, and recognise where a footnote or a more explicit paraphrase is safer than a neat but misleading one-word solution. This is not a fantasy about a machine that "knows the law" in some mysterious sense. It is an extension of what skilled legal translators already do, with access to far more text than one person can hold in mind. I am pretty sure someone is already building such an expert system.

What the evidence shows, and what it does not

This brings me back to the conditional. The philosophical argument tells us that the knowledge legal translation needs is available in language. It does not tell us that today's models use that knowledge reliably. Those are different claims. The first is about the nature of law; the second is about the state of a technology, and it must be settled empirically. The current evidence does not refute the argument, but it does not yet confirm it either.

The picture so far is promising but mixed. A Swiss legal translation benchmark found that frontier models performed strongly across laws, case summaries and press releases, while dedicated translation systems were stronger on some kinds of legal text than others. A study across four language pairs found that professional reviewers sometimes judged LLM translations as adequate and fluent as, or slightly better than, a conventional machine translation system, even when automatic scores favoured the latter. Neither result establishes parity with the best lawyer-linguists.

The failures are serious too. Research on basic legal text handling found that strong models could stumble when asked to retrieve a precise line or subsection. In a legal document, a wrong cross-reference, a dropped negation or an invented authority is not a cosmetic flaw. And a large document collection will not help if the system retrieves the wrong version of a statute, ignores the jurisdiction or confidently applies a concept from another legal tradition. More text creates the opportunity for better translation; it does not guarantee it.

“

More text creates the opportunity for better translation; it does not guarantee it.

So the best legal translation system will not simply produce fluent prose. It will make its work inspectable: name the sources it relied on, distinguish binding material from background reading, track jurisdiction and date, compare candidate renderings against both systems, and say when there is no safe equivalent. Where a translation has legal effect or rights are at stake, a qualified person must be able to review those decisions and take responsibility for the final text. That person has to come from somewhere, which means the profession must keep training legal translators even as machines take over routine work; a reviewer who never learned the craft cannot check it. Law, after all, also teaches that authority and accountability cannot be delegated to a text alone.

A domain made of language

I have argued elsewhere that machines can achieve remarkable translation without thinking as humans do. Law may be the clearest example. Its objects are constituted in language; its knowledge is recorded, cross-referenced and argued out in language; its institutions have spent centuries making meaning traceable through definitions, citations, precedents and reasons. No other domain I translate in has done so much to put its own knowledge into words.

There is still a world beyond the documents: contested facts, changing institutions, social values and the consequences of a translation for real people. And that is why I would never call law a closed system. But the legal translator's path through that world is unusually well defined by texts, and texts are what language models are made of and are extreamly capable at. If AI translation learns to follow those paths, compare the legal systems on either side and make its choices verifiable, law may prove to be its strongest domain. The next few years will confirm or deny this position.