When one writes about generative AI, there’s no escaping the feeling that Gemini, ChatGPT, or fill-in-the-blank-with-your-product-of-choice could write it up at least as well. In 2026 and beyond, so it goes. But spurred by sharing some thoughts at a faculty lunch, I thought it worth putting down how I frame recent developments as a EE who has now spent decades in the law.
So, first, what is artificial intelligence? It’s devilishly tricky to define, getting us off to a bit of a rough start. I tend to favor something like this:
A device performing a task that seems to otherwise require human intelligence.
That’s certainly going to leave boundary problems. Take the microwave’s ‘sensor reheat’—sampling humidity in order to properly heat up one’s dinner. Is that AI? Well, maybe so. A dog couldn’t do it. A tree couldn’t do it. And when a machine performs a task that would seem to require intelligence when performed by a human, we can sensibly consider it AI, even if we tend to reduce such machine behavior to ‘mere computation’ once we grow accustomed to it—a phenomenon sometimes termed the ‘AI effect.’ And even with boundary problems, this sort of AI definition can do some good work…
Artificial general intelligence or AGI, is then a device that can perform every task as well as—meaning equal to—human intelligence. When a machine can perform all tasks that humans can perform that seem to require intelligence, that will be AGI. (And I, for one, join Kent Brockman in welcoming our machine overlords.)
So defined, we’ve been using AI for a long time, and not just in our microwaves. Indeed, we’ve been using it—and using it in a mode far from the boundary cases—for decades in law.
Consider Westlaw. You might, like me, search using Boolean logic: find me every case in jurisdiction X that uses the word “confrontation” within twenty words of “face to face.” A human set on that task has a tedious job, to be sure, and it’s one that we could debate as being machine AI. (Even as a dog, gerbil, or tree certainly couldn’t do it.) Whatever that case, we all understand what is being done, and we are very comfortable with it.
But for decades—since 1992, it appears, even as I better remember the 2010 launch of Westlaw Next—there’s been an alternative: natural language search. Natural language processing (NLP) means a machine dealing in ‘ordinary English,’ or ‘ordinary Spanish,’ or whatever. Working in a human language, rather than in the languages of mathematics, including the latter’s Boolean logic. When one clicks—traditionally, I believe it was ‘More like this’—on a Westlaw headnote, what’s it do? How has West generated that content? Well, not by asking dogs, gerbils, or trees, but nor by asking human lawyers, unlike for some of its other editorial features including the headnotes themselves. Instead, like for its direct natural language queries, it leveraged extractive AI. (A semantic search, for those who know of such things.)
Extractive AI means an algorithm that searches a database in response to a natural language prompt. Could it be wrong? Could the case Westlaw claims is ‘more like this’ in actuality be nothing like this? Of course. All AI can be wrong. But we’ve lived just fine in this world. And here’s why: all humans can be wrong, too. Let’s say I instead ask my good friend Stacey, ‘Hey, Stacey, do you know any cases in the Tenth Circuit that are like this one?’ Might she answer and yet be wrong? Of course! But this is just fine. Because as much as I like Stacey, I don’t run to class with the case she cites. And I’d certainly not run to court with the case she cites. I’d read the case. Problem solved.
So, extractive AI is very familiar—it’s been in Westlaw for decades, it’s been Google for decades—and we get along just fine.
Last but not least, then, is generative AI, including its textual Large Language Models (LLMs). Generative AI is an algorithm that generates novel human-understandable content in response to a natural language prompt. It is a machine that seems to understand human language and that generates human-understandable content in return.
For example, ChatGPT means Chat Generative Pretrained Transformer. It is generating human-understandable content; it is pretrained through self-supervised machine learning; and it employs a Markov-chain-on-steroids deep neural network known as a transformer. (Again, I mention Markov chains for those who know of such things; unlike an nth-order chain, the transformer uses so-called ‘attention’ to consider all that precedes. I’m not saying it is mathematically equivalent—it definitely is not—but it harkens to something we math nerds might once have learned ‘in the day.’)
How’s it train? The model is shown billions of sequences of text and learns to predict the next word (or, technically, next token that may be a partial word) in each sequence. This requires a forward pass of its algorithm; then a calculation of how well or poorly it did, using a loss function; and then adjusting to a slightly better place through longstanding mathematical techniques of back-propagation and gradient descent. The model thereby trains to its billions of parameters.
Finally, there is generative AI that aims to be better in the legal space. The problem with extractive AI is that it will miss a relevant case. The problem with generative AI is that it will hallucinate a plausible case (just as Stacey might do). So, these legal products turn down the ‘temperature’ to favor accuracy over creativity, returning us only the most statistically likely of results, and sometimes ground the model in a limited universe of materials (so-called retrieval augmented generation).
But how’s it so damn good? Why is it that a few years ago the best models struck me as simply awful, yet contemporary products strike me as rather brilliant—where by ‘brilliant’ I mean incredibly useful. Well, nobody entirely knows. There is a beautiful 1960 paper entitled The Unreasonable Effectiveness of Mathematics, in which physicist Eugene Wigner develops how absurdly incomprehensible is the gift that mathematical language is suitable for describing reality. There is no reason, for example, that four simple, elegant equations ought to describe the reality of electromagnetics, yet Maxwell’s Equations do.
As far as I understand it, we are experiencing the unreasonable effectiveness of generative AI. In training these transformers to billions of parameters, classical intuition suggests the model would essentially memorize the training data, thus ‘overfitting.’ It would then perform wonderfully on trained text, but get novel questions wildly wrong. Yet in practice these ‘overparameterized’ systems often generalize remarkably well. Perhaps this is because human language has a mathematics we don’t consciously know… which maybe ought not surprise, since it all comes from a neural network—the gray matter in our heads.
What we know is this: contemporary generative AI is an absurdly incomprehensible gift. And if you don’t believe me, check out this great work by Adam Unikowski that has rightly received a great deal of attention, including this Bloomberg Law video.
With that introduction, here’s my thoughts on a few more points—
1. Do LLMs ‘understand’ human language? Well, that’s an impossible question… we don’t know what it means to understand human language. Might not the human neural network be, essentially, a superb stochastic transformer coupled with higher layers of what we might term ‘self awareness’ or ‘consciousness,’ in that we then reflect upon—and perhaps invent specious reasons for—why we put together the words we put together? I’d say science simply can’t answer this as of yet… it’s still the stuff of philosophical debate like Searle’s ‘Chinese Room’ problem.
2. Can generative AI outputs be novel when the model pretrains on other people’s content? Well… so do we! Everything I know was pretrained on the content of others. So, this strikes me as a nonsensical objection.
3. Can we know what a generative AI ‘thought’? How it arrived at an answer? Well… it depends upon what one means. But as to any opacity, it’s got nothing on that of a human judge. Last time I checked, we couldn’t trace the neural pathways of her brain either. But we can evaluate her output, and the same for the generative AI.
4. Aren’t these algorithms random? If they are random, they can’t be right. I mean, if I ask an associate the same question twice, I expect the same answer each time. Well… first, you ought not. Look into studies of variability in decision-making over time—influenced by everything from football team losses to meal schedules. Humans are unpredictable and inconsistent. And, more importantly, this is why generative AI turns down the temperature in the legal sphere—we want the model to return the most statistically probable path. And to the extent one thinks leveraging randomness—a stochastic process—makes a model arbitrary or capricious, that’s simply a failure to understand mathematics.
5. Can generative AI be confidential? Sure… if you contract for a zero retention system that won’t store, let alone train on, your data. But I will also say that much of this concern is silly. When the concern is tipping off the market to client confidential data, that’s not silly. But when it is, say, a law professor who worries his next idea might get stolen, it’s terribly silly. Contemporary generative AI trains on the most brilliant works humans have ever produced (and a lot of dross as well). Something tells me your next work—like mine—won’t make that better part.
6. How ‘good’ are contemporary models? Better and better every day. See, e.g., this report, and the work at Stanford HAI. My two cents? Human lawyers are on notice—you aren’t out of a job anytime soon, but you are going to need to change how you do your work.
Leave a Reply