Can large language models contribute to litigation?

Monday 27 April 2026

The legal profession has always been susceptible to the seductions of new instruments. From the typewriter to the photocopier, from electronic discovery platforms to predictive coding, each innovation has been greeted with a mixture of enthusiasm and apprehension. Today large language models—those statistical engines of text generation associated with firms such as OpenAI, Anthropic and Google—present themselves as the latest candidate for transformation of litigation practice. Their advocates promise efficiency, cost reduction and even a form of artificial reasoning. Yet the sober litigator, particularly one trained in the adversarial traditions of English law, might reasonably approach such claims with caution.

At the heart of litigation lies not merely the manipulation of language but the disciplined application of judgment. Courts do not necessarily reward verbosity; they reward precision, relevance and, above all, credibility. A pleading is not an essay, nor is a witness statement an exercise in stylistic flourish. They are instruments of persuasion grounded in fact, law and the careful calibration of risk. Large language models, for all their facility with syntax, operate upon probability distributions rather than understanding. They predict what words are likely to follow others—nothing more elevated than that. To describe this as “intelligence” risks mistaking fluency for comprehension.

This distinction becomes critical when one considers the evidential burdens inherent in litigation. A solicitor drafting particulars of claim must ensure that every allegation is capable of proof. A barrister cross-examining a witness must be alert not only to what is said, but to what is omitted, implied or evasive. These are not merely linguistic exercises; they are acts of interpretation rooted in human experience. A machine that has never stood in a courtroom, never assessed the credibility of a nervous witness, and never borne professional liability for an error cannot easily replicate such faculties.

There is moreover a structural problem in the reliance upon large language models: their tendency towards hallucination. Fabricated authorities—cases that have never been decided, statutes that do not exist—have already entered legal submissions in multiple jurisdictions. Such incidents are not trivial embarrassments; they strike at the integrity of the judicial process. A court must be able to trust that citations are real, that quotations are accurate and that representations of the law are made in good faith. The introduction of a tool that can generate plausible falsehoods with ease imposes an additional burden of verification upon practitioners—one that may erode, rather than enhance, efficiency.

Even where the output is accurate, questions arise as to provenance. Litigation often turns upon nuance: the precise wording of a contractual clause, the subtle distinction between two lines of authority, or the factual matrix surrounding a disputed transaction. If a large language model produces a draft submission, who is its true author? The solicitor who reviews it? The firm that deploys it? Or the opaque training data upon which it was constructed? These questions are not merely philosophical; they have practical consequences for professional responsibility and client confidentiality. The duty of care owed by a lawyer to his or her client cannot be delegated to an algorithm whose inner workings are neither transparent nor fully controllable.

Cost, often invoked as the decisive advantage of automation, may also prove illusory. While it is true that large language models can generate drafts at speed, the necessity of rigorous human review remains undiminished. Indeed it may be intensified. A junior associate who drafts a memorandum learns through the process of research and writing; she develops judgment through engagement with the material. If that process is supplanted by machine-generated text, the lawyer’s role risks being reduced to that of a proofreader—tasked with identifying errors in a document he did not create. The long-term effect may be a diminution of professional skill, with consequences that are difficult to reverse.

One must also consider the adversarial nature of litigation. Each party is incentivised to exploit the weaknesses of the other. If reliance upon large language models becomes widespread, so too will efforts to manipulate them—whether through the strategic framing of prompts, the introduction of misleading data, or the exploitation of known vulnerabilities. The courtroom may thus become not only a contest of legal arguments but a contest of technological manipulation, in which the party with the superior understanding of machine behaviour gains an advantage unrelated to the merits of the case.

None of this is to deny that large language models possess utility. They may assist in the organisation of large document sets, the summarisation of routine material, or the generation of preliminary drafts for internal use. In these limited functions, they resemble earlier technologies that have been successfully integrated into legal practice. But their value lies at the periphery, not at the core. The central tasks of litigation—analysis, strategy, persuasion—remain stubbornly human.

The scepticism expressed here is not born of technophobia but of professional realism. Litigation is an arena in which errors carry consequences measured not merely in inconvenience but in financial loss, reputational damage and, at times, the deprivation of liberty. It is therefore a domain in which reliability must take precedence over novelty. Large language models, for all their promise, have yet to demonstrate the consistency, transparency and accountability required to assume a central role.

The law is a human institution. It is shaped by judges, argued by advocates and experienced by those who come before the courts. Language is its medium, but judgment is its essence. Until machines can replicate not only the form of language but the substance of judgment—an achievement that remains, for now, firmly in the realm of speculation—the prudent litigator will regard large language models not as a revolution but as a tool to be used sparingly and with care.