AI can now do the unglamorous work by which senior lawyers are made. Firms that hand over all this work to Ai systems might end up saving money in the short term but realise, a decade or so later, that they have made an early drawdown on their reserves of expertise and are left with no one to promote.
In a post earlier this month, mathematician Terence Tao described a recent incident at the frontier of his field. He spoke of how, just days after a young mathematician named Stadlmann published a proof (addressing a long-standing problem about the size of the gaps between prime numbers), several artificial intelligence (AI) companies announced numerically stronger solutions. While Tao was relieved that Stadlmann managed to publish her solution before the problem was “contaminated,” he feared for the future of his discipline.
What the Answer Hides
The real value of that kind of research lies not just in finding the answer, but in the process of arriving at it—the techniques you have to invent, the dead ends you learn to recognise, and the ideas that surface along the way. This is often easy to miss, hidden as it tends to be behind the satisfaction of achieving the result. But in almost every case, the journey matters more than the destination.
AI upends this. When it hands you the proof, mathematicians have little incentive to work through it themselves. If a proof exists, all we can do is check it—not derive it afresh. As a result, the understanding that once came from painstakingly working through the problem will vanish. Tao’s worry is not that AI is wrong or its proofs sub-optimal. It is that as these systems improve, fewer people will be able to learn the way they used to.
Last week, 25 winners of the Fields Medal (Tao included), mathematics’ top honour, signed a declaration warning that the goals of AI labs were “severely misaligned” with those of mathematicians, and that this was “similar to issues that other scientific and creative professions are facing.” What they left unsaid was that the goals we set ourselves almost always have a deeper significance than the immediate reward they offer.
In professions that still train their people on the job—as professional guilds once did—the real purpose of work is not just its completion, but the skills that accrue from doing it. There is no better example of this than the legal profession.
Draining the Pipeline
As I have pointed out in earlier columns, law remains a stubbornly artisanal craft. To this day, we don’t know how to produce senior lawyers other than by putting fresh graduates through years of unglamorous work until the skills they need take hold. This is why we make them do research, carry out due diligence reviews and create first drafts of notes—because we have long believed that the practice of doing these activities is the fertile soil in which judgment germinates and grows.
These are precisely the tasks that every legal AI tool claims to be able to do far more reliably than a junior associate. Firms that buy that marketing pitch will increase their reliance on AI and hire fewer fresh associates. Those which do not may soon find themselves outpriced by those that have and will eventually be forced to follow suit. Before we know it, our pipeline for turning fresh graduates into experienced lawyers will run dry.
Many firms are reassuring themselves that junior lawyers will now learn different skills—that their role will metamorphose from doing that work to supervising the AI systems we deploy to do it. This is an overly optimistic premise that ignores the fact that people cannot meaningfully review work they themselves have never learned to do in the first place. Identifying where AI reasoning is wrong takes exactly the sort of experience and judgement that doing the work oneself builds.
Some may argue that this is simply how work evolves—that every new technology reshapes our ways of working and we will just have to adapt. I have long argued as much myself. But adapting to a new tool is not the same as giving up what the work was giving us. The task now is to shape how we use AI, keeping what matters and handing the rest to the machine.
Apprenticeship by Design
If we agree that young lawyers can acquire the essential judgment they need only by spending time on a foundational core of legal work, then we have to design our systems to ensure they always have enough of this work to gain that experience. What’s more, we need to commit to doing this even though it might be faster and cheaper to hand such work off to AI.
Figuring out what we must leave to junior associates and what we should allow AI to do will soon become one of the core decisions management makes. Firms that thoughtfully adopt AI will not ask ‘Can AI do this?’ but instead, ‘Does doing it build the judgment a young lawyer needs to become a senior?’ Whenever the answer to this question is a ‘yes,’ the task in question should be made part of a necessary apprenticeship that all young associates must go through, no matter what it might cost the firm.
It is tempting to treat junior lawyers as an ‘expense’ to automate away. But while firms that eliminate junior lawyers may save money in the short term, they are also squandering their future by drawing down a reserve of expertise they cannot replace. The use of AI in the legal profession must not be judged solely by how much work it takes off our hands, but also by how many lawyers it allows us to keep training.
This is what Terence Tao’s warning was really about. In mathematics, as in law, there are some things we should not hand to AI— not because AI cannot do them well, but because simply finding the answer was never really the point. These are tasks that serve a larger purpose, and for that reason, we should not blithely automate them away.