Law is a permanent disagreement.

Every reported case in the library is a record of at least two qualified lawyers reading the same materials and reaching opposite conclusions. One of them was instructed, in effect, to be wrong, and did not know which one they were until judgment. We have built an entire industry on that fact and then somehow persuaded ourselves that law is a body of knowledge with answers in the back.

It is not. It is a permanent, structured disagreement, and the system says so about itself at every level. Every right of appeal is the law conceding that its first answer may be wrong. Every dissenting judgment is the law preserving the argument it has just rejected, filed carefully in case the future turns out to need it. Read a Supreme Court decision and count how many judges below got it wrong on the way up; these are not fools, they are the best in the profession, disagreeing in public because the material genuinely admits of disagreement. Nothing about that is a defect in the machinery. It is the machinery.

Which is why I am unpersuaded by the framing of legal AI as a retrieval problem. Retrieval assumes there is a thing to retrieve. On the questions that actually generate fees, there is no settled answer sitting somewhere waiting to be found; there is a spread of positions of varying strength, a direction of travel that may be about to reverse, and a decision to be made about which one to stake a client's case on. I have watched the Immigration Rules rewritten under my feet more times than I can count, and some of the better results of my career came from arguing that yesterday's settled position should not survive today's facts. A system trained on yesterday's law will reproduce yesterday's law beautifully. A good part of the job is knowing when it should not be reproduced at all.

The interesting thing is that a capable model can already argue either side, and argue both well. Ask it to make the claimant's case and it will; ask it for the Secretary of State's answer and it will do that too, with equal conviction and no discomfort whatsoever. People sometimes offer this as evidence of how close the machine is getting. I read it the opposite way. The willingness to argue both sides with identical confidence is the tell. It is not sitting inside the disagreement. It is standing outside, describing the shape of it, which is a genuinely useful thing to have and not the thing a client is paying for.

Because to be inside the disagreement is to bear the consequences of choosing. That is what the codes actually regulate, and this is one of the places where the mapping is uncomfortably exact. Paragraph 2.4 of the SRA Code says you only make assertions or put forward statements, representations or submissions to the court or others which are properly arguable. Read that again with a model in the room. The rule does not ask whether the submission was generated competently, or whether it reads well, or whether something plausible produced it. It asks whether you consider it properly arguable, and the assessment is yours, personally, on your certificate. The Bar's rC9 says the same and locates it even more explicitly: your duty of honesty and integrity means you must not draft any contention which you do not consider to be properly arguable.

Properly arguable is a judgement about a disagreement. It requires somebody to have formed a view about how a tribunal is likely to receive an argument that reasonable people reject, and to be willing to be answerable for having formed it. There is no version of that duty that can be discharged by the thing that wrote the contention, because the duty attaches to consideration, and consideration is something only a party with something at stake can meaningfully do.

I find this the most clarifying of the twelve, because it explains why the gap will not close with a better model. If law were a domain with determinate answers, then a sufficiently capable system would eventually be more reliable than any of us, and the honest response would be to get out of the way. But the indeterminacy is not a limitation of our tools or our records. It is the nature of the subject. Law is made by humans, for humans, remade whenever enough humans change their minds, and resisted whenever a few of them find the courage. The disagreement never resolves, because it is not supposed to.

So the machine can state the consensus, and it can argue either side of it, and both of those are worth having on a Tuesday morning with three deadlines. What it cannot do is choose, and mean it.

Somebody has to be wrong in public, on the record, under their own name, and come back the following week and do it again. That is the practice of law. Everything else is preparation.