Contractual Deepfakes: Can Large Language Models Generate Contracts?

Notwithstanding their unprecedented ability to generate text, LLMs do not understand the meaning of words, have no sense of context and cannot reason. Their output constitutes an approximation of statistically dominant word patterns. And yet, the drafting of contracts is often presented as a typical legal task that could be facilitated by this technology. This paper seeks to put an end to such unreasonable ideas. Predicting words differs from using language in the circumstances of specific transactions and reconstituting common contractual phrases differs from reasoning about the law. LLMs seem to be able to generate generic and superficially plausible contractual documents. In the cold light of day, such documents may turn out to be useless assemblages of inconsistent provisions or contracts that are enforceable but unsuitable for a given transaction. This paper casts a shadow on the simplistic assumption that LLMs threaten the continued viability of the legal industry.

💡 Research Summary

The paper “Contractual Deepfakes: Can Large Language Models Generate Contracts?” offers a sober, technically grounded critique of the hype surrounding generative AI in contract drafting. It begins by acknowledging the impressive linguistic fluency of large language models (LLMs) – they can write stories, solve math problems, and even pass bar‑exam style questions – but stresses that this fluency is the product of statistical word‑prediction, not of semantic understanding or legal reasoning. The author introduces the notion of “contractual deepfakes”: documents that look like contracts because they reproduce familiar boilerplate, yet lack the substantive alignment with the parties’ actual agreement, the commercial context, and the applicable law.

To evaluate whether LLMs truly generate contracts, the paper first defines a “viable contract” as a document that (1) reflects the specific transaction, (2) complies with jurisdiction‑specific legal requirements, and (3) is acceptable to the parties’ counsel as meeting their clients’ needs. A contract’s enforceability is the bare minimum; true viability also demands that the clauses allocate risk appropriately, respect bargaining power, and fit the commercial purpose. The author points out that many parties sign documents they never read, which can mask the difference between a signed piece of paper and a genuine agreement.

The discussion then separates three concepts that are often conflated: (a) the abstract agreement between parties, (b) the written contractual document that records that agreement, and (c) the act of drafting the document. The paper argues that LLMs operate only at level (c): they can produce a textual artifact that resembles a contract, but they cannot create the underlying agreement nor ensure that the artifact faithfully captures it.

Technical background follows. The author explains that LLMs are trained on massive corpora to learn word embeddings and conditional probability distributions. During inference, the model samples the next token based on these learned probabilities, with limited user control over higher‑level structure or logical consistency. Consequently, LLMs excel at reproducing popular phrasing but are prone to generate internally inconsistent provisions, overlook jurisdiction‑specific mandatory clauses, and ignore the factual matrix of a transaction. The paper invokes the symbol‑grounding problem: without embodied experience, a model cannot attach meaning to words, let alone apply legal doctrines to concrete fact patterns.

The distinction between “parametric knowledge” (the statistical patterns stored in model weights) and genuine legal knowledge is emphasized. LLMs may recall that “force majeure” is often followed by a list of events, but they cannot reason whether a particular event qualifies under the governing law, nor can they balance competing interests in a way a lawyer would. The author reviews proposals to fine‑tune LLMs on legal corpora or to augment inference with external retrieval, noting that such techniques improve surface accuracy but do not solve the deeper issue of reasoning about rights and obligations.

A cost‑benefit analysis compares the effort of reviewing an LLM‑generated draft with the effort of customizing a traditional template. While LLMs can produce a first draft faster, the subsequent need for extensive human review, error correction, and contextual adaptation often outweighs any time saved. In many cases, the “draft” is so riddled with hidden contradictions that it requires a full rewrite, making the AI‑assisted workflow less efficient than conventional methods.

The paper concludes that LLMs can generate “contractual deepfakes” – textually plausible contracts – but they cannot generate contracts in the legal sense. The ability to output a document that looks like a contract does not equate to the ability to create a binding, commercially sensible agreement. Therefore, claims that LLMs will render contract lawyers obsolete or that they pose an existential threat to the legal industry are overstated. Instead, the author calls for a nuanced view: LLMs may be useful as research or summarization tools, but their role in drafting enforceable, transaction‑specific contracts remains limited and must be supplemented by substantive legal expertise.