We were doing this long before the machines.
Think about how a junior learns. Years of reading judgments, heads, pleadings, affidavits. Nobody hands you the rules for how a court writes. You soak it up. After a while you can guess what the next paragraph of a judgment will say before you turn the page. You know how an answering affidavit “should” sound. Ask a senior what a court will do with a set of facts and you get a prediction, built from every similar case they have read.
That is, roughly, what a large language model does. It has read far more than any of us, and it predicts what comes next.
The parallel goes further. Faced with a new set of facts, we relate it to what we have seen before. So does the model. And when the facts are truly new, both of us are guessing. One paper on language models and law makes exactly that point.
But the comparison breaks in places that matter.
We know when we are guessing. Or the good ones do. A decent lawyer says “I’d have to check that”. A language model’s answer sounds equally sure whether it knows or not. (Why that happens is its own post.)
We sign. Our name goes on the heads, and so do the consequences. The model has no name on the page and no reputation to lose.
We weigh values. Where a case turns on unsettled value judgments, the same paper suggests lawyers keep the edge. Patterns only take you so far when the question is what the law ought to be.
So I don’t think of Claude as a replacement brain. It is a very large, very fast language model working next to a small, slow, accountable one. The small one has to do the checking. That’s the job now.
Source: A View of How Language Models Will Transform Law (arXiv 2405.07826), on novel fact patterns and value judgments. The title of this post is my own analogy, not a term of art.