Lawyers and artificial intelligence: a cultural tension

Monday 30 March 2026

The rapid proliferation of large language models has created a curious and often underappreciated tension between two professional cultures that rarely interacted at scale before the present decade: that of the lawyer and that of the software engineer. Each is accustomed to intellectual rigour, each operates in domains where errors may be costly, and each claims a form of professional authority grounded in specialised knowledge. Yet their methods, incentives, and assumptions about work diverge so profoundly that collaboration between them, particularly in the training of large language models, is fraught with structural difficulty.

At the centre of this divergence lies the question of time. Lawyers, especially those trained in common law jurisdictions such as England or the United States, are accustomed to valuing their labour in units of billable hours. Time is not merely a metric; it is the principal commodity. The lawyer’s professional identity is tied to the careful expenditure of time in the service of precision, caution, and accountability. Advice is rendered slowly, revised meticulously, and documented exhaustively. Every clause may matter, and every ambiguity may become the seed of litigation.

By contrast, the culture of software engineering — particularly in the systems surrounding companies such as OpenAIor Google — is oriented towards scale, iteration and marginal cost reduction. Once a system is built, it may be deployed to millions at near-zero incremental expense. Engineers do not charge by the hour in the same sense; rather, they design systems whose economic value lies precisely in escaping the linear relationship between time and output. A piece of code, once written, can be executed endlessly without further human intervention.

This fundamental asymmetry produces immediate friction when lawyers are asked to contribute to the training or refinement of large language models. The process of training such models — especially in domains such as contract drafting, regulatory compliance, or litigation analysis — requires vast quantities of high-quality, annotated data. In theory lawyers are ideally placed to provide such material. In practice however their business model renders such contributions expensive, slow and difficult to scale.

A senior commercial lawyer reviewing training data does not simply label text as “correct” or “incorrect.” He or she interrogates its assumptions, considers its jurisdictional context, and reflects upon how it might be interpreted by a court. This is not annotation in the sense understood by machine learning engineers; it is a form of professional judgement honed over decades. To reduce such judgement to a scalable input for machine learning systems is to ask the lawyer to abandon, or at least compress, the very habits that define her expertise.

Moreover lawyers operate within a framework of risk aversion that is structurally incompatible with the experimental ethos of modern machine learning. The deployment of a large language model involves probabilistic outputs, continuous updating, and an acceptance that errors will occur and be corrected over time. This is anathema to legal practice. A lawyer who provides incorrect advice may face professional discipline, reputational damage, or even liability. There is no equivalent tolerance for iterative improvement in the face of client risk.

The divergence extends beyond economics and risk into questions of epistemology — that is, how each profession understands knowledge itself. Legal reasoning is inherently interpretative. It involves analogies, precedents, and the careful parsing of language within specific factual contexts. The meaning of a legal text may shift depending upon jurisdiction, court hierarchy, or even subtle differences in wording. Engineers by contrast tend to favour formalisation. They seek to reduce problems to structures that can be encoded, optimised, and generalised.

Large language models sit uneasily between these paradigms. They operate through statistical associations across vast corpora of text, capturing patterns without necessarily understanding the normative structures that lawyers consider essential. When a model produces a plausible but incorrect legal conclusion, the engineer may view this as an acceptable artefact of probabilistic modelling. The lawyer however sees a failure of reasoning that cannot be tolerated in professional practice.

The question of incentives further complicates collaboration. Law firms are not designed to produce reusable intellectual capital in the way that technology companies are. Their revenues depend upon the continued necessity of human intervention. A perfectly functioning legal language model — capable of drafting contracts, advising on regulation, and predicting litigation outcomes — would, in effect, undermine the traditional law firm business model. While individual lawyers may be curious about such technologies, their institutions often have limited incentive to accelerate their development.

By contrast technology companies are driven by the promise of scale and market dominance. For them the integration of legal expertise into language models is a means of expanding product capability and capturing new markets. This asymmetry can lead to mismatched expectations: engineers may seek rapid, standardised inputs, while lawyers insist upon bespoke, carefully contextualised contributions that resist commodification.

Confidentiality presents another structural barrier. Legal work is frequently protected by professional privilege, a doctrine deeply embedded in jurisdictions such as England and the United States. Lawyers cannot simply contribute client documents to training datasets without navigating complex ethical and legal constraints. Even anonymisation may be insufficient, as subtle factual patterns can reveal sensitive information. The result is that much of the most valuable legal data remains inaccessible for training purposes.

There is also a question of language itself. Lawyers are trained to write in a manner that is precise, often deliberately complex, and resistant to misinterpretation. Engineers on the other hand tend to favour clarity, concision and standardisation. When legal texts are used to train language models, their inherent ambiguity and contextual dependence can lead to outputs that appear authoritative but lack the necessary nuance. Bridging this gap requires not only technical solutions but also a deeper mutual understanding of each profession’s linguistic norms.

Yet despite these challenges, there are emerging pathways towards reconciliation. Some law firms have begun to experiment with alternative billing models, including fixed fees or subscription arrangements, which align more closely with the economics of software development. Similarly hybrid professionals — individuals trained in both law and computer science — are beginning to play a mediating role, translating between the cultures of the two disciplines.

In addition regulatory pressures may force a degree of convergence. As governments develop frameworks for the governance of artificial intelligence, including within the European Union, the need for legally informed model design becomes more acute. Engineers cannot simply build systems and address legal issues retrospectively; compliance must be embedded from the outset. This creates a structural demand for legal expertise that may, over time, reshape the way lawyers engage with technology.

Ultimately the difficulty of training large language models in the legal domain is not merely a technical problem. It is a cultural one. It reflects deep-seated differences in how two professions conceive of value, risk, knowledge and time. Lawyers, grounded in traditions that prize caution and individual responsibility, find themselves confronted with systems that operate at scale, embrace uncertainty and prioritise efficiency. Engineers, accustomed to rapid iteration and exponential growth, must grapple with a domain in which nuance, context and accountability cannot be easily abstracted.

The future of legal language models will depend not only upon advances in machine learning, but upon the capacity of these two cultures to understand one another. Without such understanding, collaboration will remain partial, and the promise of truly effective legal artificial intelligence will remain elusive.