Layers of lawyers and liars

As a lawyer, one of the most terri­fying arti­facts of our era is Damien Char­lotin’s data­base of AI Hallu­ci­na­tion Cases, cata­loguing instances where lawyers and pro se liti­gants have been called out for using faulty AI-gener­ated mate­rial in briefs or other docu­ments.

I’ve written previ­ously about the depre­da­tions the judi­ciary can expect from the wide­spread adop­tion of AI. But Char­lotin records new inci­dents as they happen. (And of course, just the ones that have been detected. We can suppose that many—if not most—are not.)

If you thought that inci­dents of hallu­ci­nated legal cita­tions would be reported two or three times, and then everyone would have gotten the message—nah. If anything, these inci­dents appear to be increasing in frequency with the popu­larity of LLMs.

For those just arriving: LLMs are not truth machines. Rather, LLMs assign numbers to words and then assemble those numbers into sequences that we, the human oper­a­tors, are meant to find convincing, after those numbers are converted back into the corre­sponding English words. But LLMs don’t under­stand English. LLMs don’t even know what language is. LLMs cannot inten­tion­ally make true state­ments, as that would require them to be able to distin­guish true from false. Rather, an LLM gener­ates output, which often contains sentences that happen to be true, in the same way the broken clock of proverb happens to be right twice a day. Though this compar­ison is a little unfair to broken clocks, whose failure modes are well under­stood. Not so for LLMs.

“But humans make mistakes too.” The anti­dote to a human making a mistake is another human who is better informed or more dili­gent. We have those. What’s the anti­dote to a horde of machines that generate mate­rial that doesn’t even try to be true (aka bull­shit), at a cost that will tend toward zero, at a volume that will tend toward 100% of all textual mate­rial that exists?

We’ll get back to that.

Zero-agency catastrophe

I have predicted that the AI that comes to domi­nate us—which is my idio­syn­cratic defi­n­i­tion of AGI, or arti­fi­cial general intel­li­gence—will be emer­gent, simple, brutally stupid, and conse­quen­tial but not agentic. That last part means that AGI will have effects in the real world without AGI having to “want” them to happen.

A perfect example arrived this week: a sheet-music scan­ning service called Sound­slice found that a large number of visi­tors were uploading a certain type of silly music nota­tion that the service didn’t support. Why? Because a certain commer­cial LLM was telling large numbers of people that Sound­slice supported this silly nota­tion. So what did Sound­slice do? Not wanting to disap­point these people, it imple­mented the feature. Life used to imitate art. Now it will imitate LLMs. In that sense, the risk created by an LLM arises not from what it wants, because it has no agency. It wants nothing. But humans do. Humans want the words spuming out of the LLM to be true. And humans will behave as if they are.

Much discourse on AI risk seems focused on what we might call agentic risks: how we make sure AI systems tend to produce output that supports what we want, and tend to avoid other kinds of output—a project known as align­ment. Because align­ment is diffi­cult, an AI cata­strophe arising from failure of align­ment is more likely than one arising from an AI going rogue—or even more outlandishly, “sentient”. Outlandish because an AI will never need sophis­ti­cated measures to domi­nate us, because it can passively exert unlim­ited pres­sure, like a huge rock sitting on the chest of humanity.

How easily can align­ment go wrong? Here’s a tiny yet terri­fying anec­dote from Karen Hao’s excel­lent new book Empire of AI. Below, “RLHF process” refers to “rein­force­ment learning from human feed­back”, one tech­nique for aligning an LLM according to human ratings of its output:

[L]ate one night, a researcher [at OpenAI] made an update that included a single typo in his code before leaving the RLHF process to run overnight. That typo was an impor­tant one: It was a a minus sign flipped to a plus sign that made the RLHF process work in reverse, pushing GPT-2 to generate more offen­sive content instead of less. By the next morning, the typo had wreaked its havoc, and GPT-2 was completing every single prompt with extremely lewd and sexu­ally explicit language.

Still, a key premise of align­ment is that these flaws of reasoning are internal to the AI system, and there is some way we can—even theo­ret­i­cally—control them.

But no AI system exists in isola­tion. It inter­acts with humans. Humans who are all too happy to trust and oper­a­tionalize the hallu­ci­na­tions of LLMs, thereby magni­fying their effects. Multi­plied enough times, in enough places, and we can foresee increasing risk of what I’ll call a zero-agency cata­strophe—an AI cata­strophe brought about not by malev­o­lent AI, or even failure of align­ment, but humans imputing the same level of trust to LLMs that they would to other humans, leading to wretched conse­quences.

Law’s two vulnerabilities

For that reason, law is much more vulner­able to erosion by LLMs than the sheet-music service described earlier. That company at least had a choice whether to update their product to match the words coming out of the LLM. In law, by contrast, the words on the page are the product, as the data­base of AI Hallu­ci­na­tion Cases reminds us.

This leads to what I see as two key vulner­a­bil­i­ties:

  1. Law itself. Above, I asked what anti­dote we have to the horde of bull­shit-gener­ating machines we’re putting in the hands of lawyers at basi­cally zero cost. Did you think of one? No? Then maybe you will join me in believing that LLM output, as it grows expo­nen­tially, will gum up the law in ways small and large, leaving behind toxic, untrust­worthy slop in the form of bull­shit-filled docu­ments.

    Not in elite prod­ucts like U.S. Supreme Court briefs, which will always be tended to by highly dili­gent humans. Rather, the slop will increase as you get closer to the courts of first resort, closer to the huge number of crim­inal, family, land­lord-tenant, juve­nile, and probate cases that affect the liberty and prop­erty rights of millions. In short, the costs of the slop will fall on those least able to bear it. Not for the first time.

  2. Legal prac­tice. Because they currently lose gargan­tuan sums, AI compa­nies will be selling hard to anyone who still makes money on writing. As I’ve noted else­where, the legal industry is the biggest publishing industry in the US. I expect that the tran­si­tion to AI within law offices will proceed as Hemingway described the descent into bank­ruptcy: grad­u­ally and then suddenly. The “grad­u­ally” is happening now. The “suddenly” will start when the first huge law firms start laying off docu­ment reviewers and support staff, and inhibiting their hiring of asso­ciates. This will lead to cost-struc­ture advan­tages that will in part be passed through to clients. Competitor firms will have to match these cuts or risk having these AI-powered rivals grow even larger and crowd them out.

    I’m no labor econ­o­mist, but it hasn’t stopped me before: I’m not sure big-firm part­ners have fully penciled out the lifestyle rami­fi­ca­tions of this tran­si­tion. As I under­stand it, big-firm part­ners are paid more than they bill directly. This is possible only because lower in the firm hier­archy, there are larger groups of people who are billed at a surplus to what they get paid. The money flows upward. Not for the first time.

    But as those lower tiers of labor are auto­mated away by AI, the trickle-up economics will neces­sarily fail. By contrast, in a manu­fac­turing context, the automa­tion of labor is a benefit in strictly economic terms, because the same product is getting made at lower cost. Whereas in a company built around hourly billing of human labor, AI will be devas­tating, because it makes the product disap­pear.

    “But law firms auto­mated away a lot of typists with desktop computers, and that was fine.” I didn’t run a big law firm in the 1980s. But I’m skep­tical that analogy is econom­i­cally apt. Law firms bake many support costs into the rates of their bill­able staff. If those support costs can be auto­mated, then of course it’s bene­fi­cial. Costs always want to be mini­mized. With AI, however, the automa­tion replaces the bill­able services. So it’s a cost shrinker, but also a revenue shrinker.

    Against that back­drop, I expect that a rising wave of AI tools targeted at legal prac­tice will even­tu­ally nudge many firms to either a) consciously avoid AI, and collec­tively accept lower profits per partner in line with their firm’s less compet­i­tive economics; or b) accept minimum AI, and dismiss just enough equity part­ners to keep the profits per partner stable; or c) embrace AI, and cut deeper into the partner ranks to concen­trate the returns from AI around a smaller group, thereby increasing profits per partner. You don’t need an LLM to guess which of these paths are most likely.