Two firms receive the same case file. One knows what page 2,044 says. The other hopes it doesn’t matter. The difference between them isn’t going to be in the talent of their lawyers, and this year it’s increasingly going to be somewhere else: in whether they understood what they bought, or whether they bought a word.

The word doesn’t mean what it sounds like it means

“Artificial intelligence” doesn’t name a technology. It names a goal: building machines capable of doing things that, if done by a person, we’d call intelligent. It’s the name of the entire field and of its ambition — which is why it’s less a technical category than a promise.

That’s why “this tool uses artificial intelligence” isn’t a claim you can evaluate. It’s the same kind of statement as saying a firm “has experience.” It’s not false; it just doesn’t tell you anything. The questions that actually decide things come after.

Machine learning is a concrete method. Instead of programming rules one by one, the system is shown enormous quantities of examples and derives a pattern from them. You don’t teach it the rule: you give it the material the rule is drawn from. It’s one path toward that goal, and today it’s the dominant one, to the point that practically everything you’ll be offered under the label “AI” is, underneath, machine learning.

So the relationship is simple: artificial intelligence is the destination, and machine learning is the vehicle almost everyone is riding in today. When a vendor tells you the first, they’re telling you where they’re pointed. They haven’t yet told you what they built.

Three words and you’re set

Training, model, inference. Three words. With them you can read any pitch that lands on your desk this year.

Training is the process: enormous amounts of text get processed and a system gets tuned until it predicts well. It’s expensive, it’s slow, and it happens once, at the vendor’s shop, long before you ever see the product.

Model is what’s left over from that process: a file, a finished and frozen object. It doesn’t learn from you while you talk, and it doesn’t show up tomorrow knowing more than it did today because it dealt with you.

Inference is every single time that model gets used: your question goes in, an answer comes out. And here’s what no demo tells you: everything an AI system answers is an inference. It isn’t a lookup against a database of truths, and it isn’t human reasoning. It’s a prediction — made with an enormous amount of information and an enormous amount of statistics — about what text comes next.

A language model is a model trained specifically on text, and it’s the piece underneath almost everything you’ve already tried. ChatGPT, Claude, Gemini, and the “AI legal tools” you’ve been offered aren’t models: they’re applications built on top of a model, each with its own interface, its own permissions, and its own price. When you compare two of them, you’re often comparing two wrappers around the same engine.

Two consequences you end up signing

If every answer is a prediction, two things follow. Neither one is a rare defect a patch is going to fix next month.

The first is that the system can state, with total ease, something that never happened. It’s called a hallucination. It isn’t an occasional glitch: it’s the structural consequence of the mechanism. A system trained to produce the most plausible text produces plausible text even when it has nothing to back it up, and plausible text — well-written, stated with confidence, without a single hedge — is exactly what a professional pressed for time accepts without checking. Fluency isn’t evidence of accuracy. Fluency is what the system has guaranteed; accuracy isn’t.

The second is that the model keeps no memory and doesn’t hold your entire case file in view — the technical reason the tool you paid for “forgot” half the document (why none of those tools worked for you).

The bottleneck was never intelligence

It’s worth stopping here, because this is where the market is arguing about the wrong thing.

The problem for a litigator facing a large case file isn’t a lack of intelligence. It’s that the file doesn’t fit in the time available. Deep reading moves at roughly 44 to 100 pages an hour depending on the type of document; a case file of tens of thousands of pages is, therefore, somewhere between five hundred and eleven hundred hours of careful reading. Three to six person-months full time. Even at a modest reviewer’s rate, that’s on the order of tens of thousands of dollars of human labor for a single complete read.

But the cost isn’t the serious part. The serious part is that a lone litigator simply can’t do it. That’s why something always “slips through.” And in criminal matters, what slips through isn’t an administrative detail: an omitted mention, a contradiction nobody cross-checked, can change the outcome of a case where someone’s liberty and due process are at stake.

That’s where this technology is worth something, and not before. It doesn’t replace a lawyer’s judgment: it frees it up. It takes the thousand hours of mechanical reading off your hands and hands you back a case file that’s already been gone through, so you spend your judgment on the one thing that can’t be delegated: deciding what it means.

It’s the same relationship you already have with an expert report, or with the draft a junior associate hands you. The expert doesn’t decide the case. The junior doesn’t sign the brief. Both hand you worked material that you review, correct, and make your own, and neither one saves you the review. With an artificial intelligence system the rule is identical, and stricter still: nothing gets signed without checking it against the source. A claim that can’t be traced to the exact page it came from isn’t a finding. It’s a well-written hypothesis.

The question you actually can evaluate

Now go back to the pitch on your desk, three words in hand. Don’t ask whether the tool “uses artificial intelligence”: you already know that’s not evaluable. Ask what it does with your entire case file, and ask whether every claim it produces can be opened at the page it came from.

The second one is the one that decides. And you’d better ask it soon, because artificial intelligence isn’t going to take the case from you: the colleague who actually read the whole case file is.