Artificial intelligence is changing legal practice at remarkable speed. Lawyers can now use generative AI to summarize hundreds of pages of disclosure, prepare first drafts of facta and correspondence, analyze authorities, identify issues and refine arguments in a fraction of the time those tasks once required.
Much of the discussion surrounding this transformation has focused on an obvious question: Will AI eventually replace lawyers? That question may be less important than another possibility. What if AI does not replace lawyers at all, but instead causes legal practice to reorganize gradually around the technology until some of the human skills and institutional practices upon which law depends become difficult to recover?
That is the provocative argument advanced by Joseph J. Avery in his article, When We Stop Making Law: AI Drift in Legal Practice. Avery is not predicting that judges will suddenly surrender decision-making to machines or that lawyers will disappear from courtrooms. His argument is considerably more subtle. Technologies rarely eliminate established practices by decree. Instead, institutions adapt incrementally to technological convenience until practices that were once optional become effectively indispensable.
The Air Conditioning Problem
Avery illustrates the problem with an unexpectedly useful analogy: air conditioning.
Air conditioning initially supplemented the way buildings functioned. Buildings could operate without it because windows opened and architecture incorporated natural ventilation, shade and other forms of passive cooling. As mechanical cooling became ubiquitous, however, buildings themselves changed. Architects increasingly designed structures on the assumption that air conditioning would always be available. Windows stopped opening and natural ventilation became less important.
Eventually, a technology that had once supplemented the building became essential to its operation. Nobody made a conscious collective decision to eliminate naturally ventilated buildings. The surrounding environment simply evolved until the previous way of doing things became impractical.
Avery argues that law may be approaching a similar transition. AI begins as an assistant that summarizes cases, organizes disclosure or produces a first draft. Its role then expands. It drafts pleadings and facta, compares evidence, identifies inconsistencies and assists judges in reviewing submissions. Courts adapt procedures to increased speed and volume. Clients expect work to be completed more quickly. Firms adjust staffing and billing models. Law schools train students to use AI from the beginning of their legal education.
At some point, practicing law without AI could become analogous to operating a modern office tower without air conditioning: theoretically possible, but practically and economically unrealistic. The difficult question is what happens to the human elements of legal practice that may be displaced during that transition.
Law Is More Than Producing the Correct Answer
Lawyers understandably tend to evaluate technology by reference to accuracy. Did the AI identify the correct authority? Did it accurately summarize the judgment? Did it identify the governing legal test? Did it produce a persuasive argument?
These are important questions, particularly given the documented problems associated with hallucinated authorities and inaccurate AI-generated legal research. Accuracy, however, does not fully capture what the legal system is designed to accomplish.
Law is not merely a mechanism for generating technically correct answers. It is also a process through which human beings deliberate about fairness, responsibility, credibility, reasonableness and justice. Many of those concepts resist mechanical application because the legal rules themselves deliberately leave room for judgment.
Consider the familiar family-law concept of the best interests of a child. Legislation identifies factors, appellate decisions establish principles and trial decisions provide guidance, but there is no formula capable of producing an objectively correct parenting arrangement for every family. Parents give evidence, lawyers advance competing interpretations of that evidence, judges assess credibility and context, competing interests are weighed and reasons are ultimately given.
Those processes are not merely inefficient steps standing between a dispute and its correct answer. They contribute to the legitimacy of the answer itself. Law derives part of its authority from the fact that affected individuals have been heard, their arguments have been considered and a person exercising lawful authority has accepted responsibility for the resulting decision.
Why Family Law Makes the Problem Particularly Clear
Family law provides a particularly useful illustration because its disputes frequently require contextual and intensely human judgment.
Imagine two separated parents disputing whether their nine-year-old child should relocate from Toronto to Vancouver. A sufficiently sophisticated AI system could conceivably review Canadian relocation jurisprudence, analyze the statutory best-interests factors, assess the evidence and calculate that one parent has a 68 per cent probability of succeeding.
That information could be extraordinarily valuable to the lawyers and their clients. It could inform settlement discussions, identify weaknesses in the evidence and help the parties understand litigation risk. The more difficult question is whether the same system should ultimately decide the case.
Most participants in the justice system would instinctively resist that proposition, and the reasons matter. We expect a person exercising state authority over a family to engage meaningfully with the people affected by the decision. We expect the judge to hear the evidence, consider the competing arguments, appreciate what is at stake and provide reasons explaining why the family’s life is about to change.
Avery points to empirical research suggesting that perceptions of legal authority can decline when people believe artificial intelligence participated in judicial reasoning, even where the ultimate outcome remains unchanged. If that research proves durable, it suggests that the legitimacy of adjudication depends not simply upon the quality of the outcome but also upon the process by which the outcome was reached.
For lawyers and judges experimenting with generative AI, that distinction deserves serious attention.
The Efficiency Paradox
There is also reason to question one of the central assumptions underlying legal AI: that greater efficiency will necessarily reduce the amount of legal work being performed.
The opposite may occur.
If AI allows counsel to produce a 40-page submission in the time previously required to prepare 15 pages, opposing counsel must now review and respond to 40 pages. They also use AI, producing a 50-page response. The judge is then confronted with 90 pages of submissions and uses AI to summarize them.
Clients can participate in the same escalation. Generative AI makes it possible for a client to produce lengthy chronologies, draft arguments, correspondence and document analyses at virtually no marginal cost. Lawyers who once had to organize scarce information may increasingly find themselves responsible for filtering an abundance of it.
The result could be an efficiency paradox in which AI reduces the cost of producing legal material while simultaneously increasing the total amount of material that participants in the justice system are expected to process. AI then becomes indispensable to managing the volume that AI itself helped create.
In that environment, opting out becomes increasingly difficult. A lawyer who refuses to use AI may eventually resemble the lawyer who refused to adopt email: perhaps principled, but increasingly unable to compete within an ecosystem designed around the technology.
The Deeper Risk Is Deskilling
Much of the professional discussion about AI has understandably concentrated on immediate risks such as hallucinated authorities, confidentiality, privilege and inaccurate output. Those concerns are serious, but they are also visible and therefore potentially manageable through professional standards, technological safeguards and careful human verification.
Deskilling may prove more difficult because it occurs gradually.
Legal judgment develops through repetition and intellectual struggle. Junior lawyers learn research partly by getting lost in cases before discovering the authority that matters. They learn drafting by struggling to express an argument clearly. They learn advocacy by constructing arguments that fail and understanding why. Perhaps most importantly, they develop judgment by watching experienced lawyers decide which arguments should not be made.
Anyone who has trained junior lawyers understands that the document produced at the end of an assignment is only part of its value. The intellectual process required to produce the document is often where professional development occurs.
Generative AI potentially changes that process. If a new generation of lawyers routinely begins with an AI-generated analysis, research summary and draft argument, their finished work may be better and produced considerably faster. It does not necessarily follow that they are developing better legal judgment.
GPS provides an obvious analogy. It gets us to our destination more efficiently while potentially weakening our ability to navigate without it. AI could produce a similar effect in legal practice. The long-term danger is not that the technology becomes extraordinarily capable. It is that lawyers gradually become less capable of practicing without it.
AI Also Challenges the Economics of Legal Practice
Generative AI creates an equally uncomfortable challenge to the traditional relationship between time and value in legal services.
Suppose legal research that once required an associate four hours can now be generated, reviewed and independently verified in 45 minutes. What should the client pay for: four hours, 45 minutes, the value of the answer, or some combination of technology, expertise and professional responsibility?
The profession has been able to postpone answering that question because hourly billing remains deeply embedded in law-firm economics. AI will make the contradiction increasingly difficult to ignore. Clients will eventually understand that many tasks historically requiring several hours can now be completed much faster.
That does not necessarily make lawyers less valuable. It may have the opposite effect. If routine intellectual production becomes inexpensive, the distinctly human components of legal representation — judgment, strategy, negotiation, advocacy, counselling and accountability — may become more valuable.
Law firms will nevertheless have to become much clearer about what clients are purchasing. If clients are no longer principally paying for the time required to produce information, firms will need to articulate more convincingly the value of the professional judgment applied to that information.
The Lawyer Still Owns the Work
Whatever technological changes occur, one principle should remain straightforward: the lawyer remains professionally responsible for the work.
If an AI system invents an authority and counsel files it, responsibility rests with counsel. If the system misunderstands evidence and the lawyer adopts its analysis, the lawyer remains accountable. If confidential client information is improperly disclosed through an AI platform, the existence of the technology does not relieve the lawyer of professional obligations.
The same is true of subtler errors. An AI-generated argument may be legally plausible and rhetorically impressive while fundamentally misunderstanding the client’s objectives, the dynamics between the parties or the strategic consequences of advancing a particular position.
Generative AI can produce language and analysis. It cannot assume professional responsibility for the decision to use them. That distinction may become one of the defining characteristics of legal practice in the AI era.
What Should Law Firms Refuse to Automate?
The practical question for law firms is therefore no longer whether they should use artificial intelligence. AI will be part of legal practice, and firms that refuse to engage with it are unlikely to avoid the broader institutional changes it produces.
The better question is which functions should be automated aggressively and which should remain substantially human.
Research assistance, document organization, chronologies, routine drafting and administrative processes are obvious candidates for substantial automation. There is little virtue in preserving repetitive work merely because lawyers have traditionally performed it manually.
Judgment, client counselling, credibility assessment, negotiation, strategic decision-making and professional responsibility are different. Those functions lie much closer to the reasons clients retain lawyers and to the institutional foundations upon which legal authority depends.
The objective should therefore not be to preserve inefficient legal work. Nor should the profession romanticize endless hours of manual research and drafting that technology can perform more effectively. The challenge is to distinguish between legal labour and legal judgment, using AI aggressively to reduce the former without inadvertently surrendering the latter.
We May Not Notice the Line Until We Have Crossed It
Avery’s most important warning concerns the way technological change actually occurs. There is unlikely to be a meeting at which the legal profession collectively decides that human deliberation is no longer necessary. Instead, the transformation will consist of thousands of individually reasonable decisions.
Let AI summarize the case. Let it prepare the chronology. Let it compare the affidavits. Let it draft this section of the factum. Let it assess the competing positions. Let it recommend an outcome. Eventually, perhaps, let it prepare the first draft of the reasons.
Each decision may be defensible in isolation. The risk lies in their cumulative effect.
The profession could eventually find itself in a world where lawyers routinely review arguments they did not formulate, judges approve reasons they did not meaningfully construct and clients receive advice produced through analytical processes that neither lawyer nor client fully understands. The resulting documents may look exactly like law, and the outcomes may even be correct.
Yet something important could nevertheless have disappeared.
The central challenge presented by artificial intelligence is therefore not whether machines can learn to perform the work lawyers currently perform. They almost certainly will become remarkably capable of doing so.
The more consequential question is whether, after years of allowing machines to perform an increasing share of the intellectual work of legal practice, lawyers will still know how to think like lawyers.
