A.I.’s dawn is now daylight

Note: it seems apt that this article has been written, at least partially, using A.I.

A.I. has actually been around far longer than most people think. When we refer to A.I. what we are most often referring to in this article is a Large Language Model (“LLM”). To many, it looks a little like a very sophisticated search engine; indeed, in many respects, it is!

In the early 1990s IBM used statistical models to pioneer word-alignment techniques for machine translation. However, the significant shift did not happen until 2017, when transformer architecture started being used. The science behind this transformer architecture is, as you would imagine, quite complex, but it enables the model to process the relationships between all elements of a sequence simultaneously, irrespective of the distance between those elements.

Anyone who understands even basic computer programming will realise that this structure places a great deal of demand on the computer processing the query, particularly where the query becomes long and technical. Despite this resource draw, A.I. has profoundly impacted almost every aspect of the workplace in ways many never thought would happen.

Hardly a day goes by where you do not hear of a lawyer using a “hallucinated” case in court, or where a judge admits to using A.I. to write, or edit, a judgment.

Yet beneath the headlines, the reality is rather more nuanced. Much of the public discussion surrounding A.I. in the legal profession tends to oscillate between two extremes: either A.I. will replace lawyers entirely, or it is little more than an unreliable novelty machine incapable of serious professional use. Neither proposition is correct.

The legal profession has, historically, been slow to adapt to technological change. Indeed, many chambers and firms only fully embraced remote working technology because they were effectively forced to do so during the COVID-19 pandemic. Before that, practices which are now routine such as electronic bundles, digital disclosure exercises, virtual hearings and cloud-based document management were often regarded with suspicion by practitioners who had spent decades relying on paper files and physical attendance.

A.I. represents another such inflection point, albeit on a far greater scale.

Used properly, LLMs are capable of dramatically increasing efficiency in areas which have traditionally consumed substantial fee-earner time but relatively little intellectual creativity. The review of lengthy correspondence, first-pass document summarisation, chronology creation, proofreading, drafting assistance and legal research can all, to varying degrees, be accelerated through the sensible deployment of A.I.

A junior solicitor who may once have spent six hours constructing a chronology from several lever-arch files of disclosure may now be able to produce a workable first draft in under an hour. Likewise, a barrister preparing for a conference can use A.I. to rapidly distil large quantities of background material into a manageable summary before applying their own legal analysis and strategic judgment.

The important distinction, however, is that A.I. is presently most effective as an assistant rather than as a decision-maker and that is where it has been used improperly.

That distinction matters enormously in law because legal practice is not merely the retrieval of information. It involves judgment, proportionality, ethics, strategy and an appreciation of nuance which frequently extends beyond the written word. A witness statement may be technically coherent yet strategically disastrous. A point of law may be legally arguable yet commercially absurd. A settlement proposal may be objectively generous but psychologically incapable of acceptance because of the personalities involved. These are fundamentally human considerations and A.I. as yet cannot fully understand these nuances.

This is where some of the more sensational predictions about the demise of the legal profession begin to unravel. Clients rarely pay lawyers simply to locate information. They pay for judgment under conditions of uncertainty.

That is not to say the profession will remain untouched. Quite the opposite. A.I. is likely to place substantial pressure on the traditional economic structure of legal services. Many forms of junior-level work which once justified significant chargeable hours (ie, disclosure) may become increasingly difficult to bill at historic rates where clients know such work can be materially accelerated through technology.

This creates an uncomfortable question for the profession: if A.I. can reduce a ten-hour task to a two-hour task, is the client paying for the lawyer’s time, or for the lawyer’s expertise?

Historically, many legal billing models have depended heavily upon time expenditure rather than value creation. A.I. may accelerate the transition away from that model. Clients are unlikely to tolerate being charged for large volumes of administrative or repetitive work where technological tools substantially reduce the labour involved.

Equally, however, A.I. introduces risks which are potentially profound.

The most obvious of these is the phenomenon now commonly referred to as “hallucination”. Unlike a conventional search engine, an LLM does not “know” whether a proposition is true. Many users of ChatGPT will have experienced the fickle nature of its responses largely dependent on how you ask the question. Rather, it predicts what combination of words is statistically likely to follow the prompt it has been given. Most of the time this produces remarkably convincing results. Occasionally, it produces complete fiction with equal confidence.

For lawyers, that presents a uniquely dangerous problem.

The legal profession operates upon authority, precision and trust. A fabricated case citation in a skeleton argument is not merely embarrassing; it potentially amounts to professional misconduct. Courts across multiple jurisdictions have already encountered submissions containing non-existent authorities generated by A.I. systems. In many of those cases, the issue was not simply that the lawyer used A.I., but that the lawyer failed to verify the output before deploying it professionally.

In truth, this is perhaps less a technological failure than a human one.

A competent lawyer should no more blindly rely upon A.I. output than they should blindly rely upon an unverified statement from a trainee, a witness, or indeed another solicitor. The obligation to exercise professional judgment remains non-delegable.

There are also subtler dangers emerging.

One concerns confidentiality and privilege. Many publicly accessible A.I. systems operate by processing user inputs through external servers. Practitioners who upload privileged documents, sensitive client information or commercially confidential material without understanding how the platform stores or uses that data may inadvertently create serious confidentiality and data-protection issues.

Another concern lies in over-reliance. The more fluent and persuasive A.I. becomes, the greater the temptation to accept its conclusions uncritically. Yet legal analysis often depends precisely upon identifying the unusual case, the hidden factual distinction or the exception to the apparent rule. A.I., by its nature, tends toward consensus prediction. Law, however, is frequently won in the margins.

There is also a broader jurisprudential question developing beneath the surface: what happens when access to competent legal drafting becomes democratised?

Historically, the ability to produce structured legal correspondence, coherent pleadings or technically competent submissions required either legal training or the financial means to instruct someone who possessed it. A.I. lowers that barrier dramatically. Litigants in person are increasingly able to produce documents which, superficially at least, resemble professionally drafted material.

That development may improve access to justice in some respects. It may equally place additional strain upon courts already burdened by increasing volumes of litigation. Judges may find themselves dealing with a growing number of technically polished but substantively flawed claims generated through automated systems that give the appearance of legal merit without the underlying analytical foundation. In only April of this year, the Bar Standards Board noted a 25% increase in complaints, largely driven by A.I. drafted complaints.

Similarly, warnings and terms are entering into Court and arbitration orders in respect of the use of A.I. in the legal process requiring parties to identify whether they have used A.I. and to what extent. The Commercial Court Guide for instance has long stated in respect of Trial Witness Statements that they:

“must set out only matters of fact of which the witness has personal knowledge that are relevant to the case, and must identify by list what documents, if any, the witness has referred to or been referred to for the purpose of providing the evidence set out in their trial witness statement. […]”

And:

“Factual witnesses give evidence at trials to provide the court with testimony as to matters of which they have personal knowledge, including their recollection of matters they witnessed personally, where such testimony is relevant to issues of fact to be determined at trial”

And also:

“The duty of factual witnesses is to give the court an honest account of matters known personally to them (including, if relevant to the issues in the case, what they recall as to matters witnessed personally by them or what they would or would not have done or thought if the facts, or their understanding of them, had been different). It is improper to put pressure of any kind on a witness to give anything other than their own account, to the best of their ability and recollection, of the matters about which the witness is asked to give evidence.”

Where does the above leave witnesses relying heavily upon A.I. to generate their witness statements? Can it be said to be their evidence if A.I. is creating novel interpretations of particular documents rather than the witness themselves?

The legal profession therefore finds itself in an unusual position. A.I. is neither an existential threat nor a passing fad. It is a tool of immense capability which will reward practitioners who understand both its strengths and its limitations.

The lawyers most vulnerable to A.I. are unlikely to be those exercising sophisticated judgment, advocacy and strategic thinking. Rather, they are those whose practices depend heavily upon routine information processing capable of automation.

The more interesting question is not whether A.I. will replace lawyers. It is whether lawyers using A.I. will replace lawyers who refuse to use it at all.

The difficulty for the legal profession is not that A.I. exists; it is that it exists in a form which appears deceptively competent. Unlike earlier software tools, which were plainly mechanical and constrained, modern LLMs communicate with remarkable fluency and confidence. That confidence, however, is entirely synthetic. An LLM does not “know” anything in the conventional sense. It predicts language patterns based upon probabilities derived from unimaginably large datasets. In simple terms, it is designed to generate the most statistically likely next word in a sequence. The fact that this process can produce cogent legal analysis is both extraordinary and, potentially, dangerous.

For lawyers, the temptation is obvious. The legal profession is built upon language, structure, precedent, synthesis and analysis — all areas in which LLMs appear to excel. Tasks which previously consumed hours can now be completed in minutes. Draft witness statements, first-pass advices, chronologies, disclosure summaries, research notes and even skeleton arguments can be produced almost instantaneously.

Used properly, this technology is capable of dramatically improving efficiency and reducing cost. Junior lawyers can analyse larger volumes of material in shorter periods. Barristers can refine arguments more rapidly. In-house legal departments can process routine contractual queries without requiring external advice on every occasion. In litigation, document-heavy exercises which once required armies of trainees may increasingly become matters of technological management rather than manpower.

There are also significant access to justice implications. It is difficult to ignore the possibility that A.I. may, in time, assist individuals who could never otherwise afford legal advice. Small businesses may obtain preliminary contractual guidance at negligible cost. Litigants in person may be able to better understand procedure and legal concepts which were previously inaccessible to non-lawyers. Courts themselves may ultimately utilise A.I. to assist with case management, document organisation and the summarisation of evidence.

However, these benefits only materialise where the technology is used with discipline and scepticism. The greatest danger posed by LLMs is not that they occasionally make mistakes; lawyers make mistakes too. The danger is that they make mistakes persuasively.

A hallucinated authority is not produced hesitantly or ambiguously. It is often presented with complete confidence, accompanied by fabricated citations, procedural histories and invented quotations. To an inexperienced practitioner, or to one working under significant time pressure, the output may appear entirely legitimate. Indeed, one of the more troubling aspects of A.I.-generated legal work is that the better the prose becomes, the more difficult it is to identify where the analysis fails.

This issue has already manifested itself repeatedly before the courts. Judges across multiple jurisdictions have criticised practitioners for relying upon non-existent authorities generated by A.I. systems. In some cases, entirely fictitious judgments have found their way into pleadings and written submissions. The reputational consequences for the lawyers involved have been severe, but the broader concern is systemic. The administration of justice depends fundamentally upon the integrity and reliability of legal authorities. A profession which begins to outsource verification, risks undermining the very foundation upon which precedent-based systems operate.

Confidentiality and privilege presents a further difficulty. Many publicly available A.I. tools operate by processing user prompts through external servers, often subject to opaque data retention policies. Lawyers who upload pleadings, advices, witness evidence or commercially sensitive documents into unsecured systems may inadvertently expose privileged or confidential information. In some respects, the legal profession’s enthusiasm for generative A.I. has moved faster than its understanding of the associated data protection and confidentiality risks.

There is also an emerging professional competence issue. Historically, junior lawyers developed their skills through repetition: reviewing authorities, drafting documents, analysing disclosure and learning, often painstakingly, how legal arguments are constructed. If those foundational tasks become increasingly automated, there is a legitimate question as to how future lawyers acquire the judgment necessary to supervise the technology itself. Put differently, if a generation of lawyers becomes accustomed to accepting machine-generated analysis without fully understanding the underlying reasoning, the profession may ultimately weaken rather than strengthen.

None of this is to suggest that A.I. should be resisted. That would be both unrealistic and commercially self-defeating. The firms and practitioners who refuse to engage with these tools will almost certainly find themselves outpaced by those who use them intelligently. The more realistic question is not whether A.I. should be used, but how it should be governed.

At present, the profession appears to be moving toward a model in which A.I. is treated as an assistant rather than a substitute. That distinction matters. Used correctly, an LLM can be an extraordinarily powerful drafting and analytical aid. Used carelessly, it can become a mechanism for amplifying error at unprecedented speed. The responsibility for the final product, however sophisticated the software becomes, will continue to rest with the lawyer whose name appears upon the document. Similarly in answer to the question above about whose evidence is it if A.I. writes it? This will again come down to the person whose name appears at the bottom of the witness statement. Use A.I. as a witness and you may find your case seriously harmed and your evidence thrown out.

Perhaps that is the central paradox of A.I. within the legal profession. The technology is at its most useful when deployed by lawyers who are already highly skilled, because only experienced practitioners possess the judgment necessary to interrogate its output properly. In that sense, A.I. may not replace good lawyers at all. It may simply expose bad ones more quickly.

Kyle Lecuona

Barrister

London

The above article does not constitute legal advice. Every dispute is unique and expert advice should be sought in respect of your specific circumstances.