AI & Law Stuff
#22 Smarts legibility, Large Libel Models, and Consulting Mishaps
No one Sees you Fail
How do you pick a lawyer ?
There are several possible answers (one of which, we saw, is personality), but it’s a fair bet to say most people would immediately go for competence: you want a good lawyer, if not the best, a lawyer who’s got smarts, who will take you through your case and back and help you with your legal pickle. To the extent that law is something that is contestable, you want the person able to win that contest, and in an intellectual profession, that means a smart lawyer. Setting aside purchasing power for a moment, all things being equal, people don’t compromise on quality in this respect.
Yet, this pushes the question to: “how do you know ?” Which is where the theory meets (and crashes) with how we actually pick lawyers, since competence and smarts are hard to gauge. Reputation certainly helps, but everyone can cite people famous despite terrible flaws. Word of mouth, the grapevine, is how most of us orient ourselves in picking, and when that’s done, familiarity, recommendation, family and student links are the main ways to discover legal talent that can serve us.
This works because competence is not readily legible. There is no actual scoreboard of the “best lawyer” in concrete cases, or at least not any that manages to feel other than the legitimate marketing exercise of the various rankings, top 50, etc. Lawyers tend to resist such scoreboards for several reasons, some legitimate,1 others less so. In the meantime, although some use degrees, experience, brand names, or prices to sort themselves out and appeal to our prejudices, others can readily invoke the rather appealing counter-argument that all this means nothing, that the “real” skills of lawyering exist elsewhere, that underdogs deserve the win, etc.
Now reintroduce prices, and you see that the legal market relies on this lack of legibility. If performance were legible, palpable, something you could get a feel for very quickly, nobody would find it fair that “who gets the best” relies entirely on “who can afford the best”. Instead, we can all believe that we picked the “best” lawyer (all things considered) and feel no particular opportunity cost at not switching to someone else. The lack of legibility protects not only the distribution, but also the size of the market.2 The comfort of not knowing whether your lawyer is the best is also the comfort of never having to find out she was not.
Anyhow, we all ran an experiment in that vein last week, after Anthropic was made to suddenly remove access to Fable, their most advanced available model. Pieter Garicano, writing for Silicon Continent this week, put words to what I assume is a common feeling:
Fable teaches us something general about the ‘intelligence commoditization’ thesis and the ability of worse models (whether Chinese or European) to substitute for the frontier ones. Having experienced a better model, Opus 4.8 is painful to use even on tasks where it previously felt helpful.
Which comes down to my point above: when it comes to smarts, people don’t easily settle for second best.
Now, that’s true if AI (or a lawyer, for that matter) is used for tasks that are open-ended and hard to evaluate. This is what distinguishes smarts from mere tools: the latter is worth acquiring if it does the job (say, drive a screw), and once that job is ensured, people can opt for their own trade-off between price, durability, branding, etc. Legal work, however, does not always match that description and for everything else, smarts trump it all.
As well, this puts into perspective the (evergreen) news that some legal tech providers are training or fine-tuning their own models.3 Past research (including a very recent Nature paper in the same line) has not been kind on the merits of fine-tuned models compared to more general models. Hence some reasons to be skeptical, until, that is, the models are great at tackling some distinct tasks whose outcome can be easily assessed.
The fine-tuners, just as some for the “AI-first law firms”, are betting many legal tasks are like a screw to drive, but it’s on them to prove that there is a demand for their supply. Until then, people will likely continue to go for the best, and rely on the old heuristics to that end.
[Programming Note] I have enabled pledges for Artificial Authority. Regular posts will always be free ; this is for those who find the newsletter or the hallucination database useful and want to support my work.
Large Libel Models
LLMs by nature hallucinate, (nearly) all of us know, and these hallucinations have consequences. One frequent scenario is seeing an LLM or AI output say things about you that aren’t true and could qualify as defamatory.4 We are still at an early stage in policing this potential liability region, and there is a useful tracker from Ed Lee cataloguing these (and other torts) lawsuits in the USA.
From first principles, the judicial answer will likely depend on the exact scenario that led to the alleged libel. Some complaints rely on what is possibly the crudest form: asking an LLM point blank about oneself, and seeing it spill out misconceptions.5 This kind of direct elicitation sounds hard to blame on the AI company, but maybe I am wrong and courts will disagree. A thornier scenario, however, involves AI outputs available in the wild about you - something you did not elicit but anyone else might stumble upon (e.g.).
Seemingly, this latter scenario has now led to what looks, at first, like conflicting jurisprudence from German courts. In one case, a Munich court reportedly found (judgment here, in German) that AI-generated outputs from Google that had associated some German publishers with shady practices were libellous and engaged the firm’s responsibility. In a second case, a Berlin court reportedly held that such AI overviews could not engage Google’s liability in a trademark infringement case.
At stake in both cases are complex and unsettled questions about the extent of control Google has on the AI overview outputs, and that question showcases a tension: if these overviews are produced at scale and remain open-ended, there is no way to control them except ex ante through the production methods. But those, relying ultimately on LLMs and their probabilistic limitations, will make it hard to get it right 100%. In other words, you’ll always find a potential aggrieved victim of AI outputs.
Given this, there are three models of what happens if liability findings keep piling up, and all the more so if they answer cruder types of self-generated libels. Down the line, either:
Model providers find a fix, but good luck with that; or
Some AI functionalities stop functioning, or filters are put in place to prevent AI dealing with natural persons the same way they can’t teach you to build a nuclear bomb; or
Jurisprudence gives way somehow.
Now, it’s fair, and likely good, that courts tasked with resolving a single case do not think through it in terms of equilibrium and down-the-line technical consequences. Fiat iustitia et pereat mundus, judges might think, and hope for the best as everyone adapts to what’s coming.
But that’s what is interesting with these developments: AI, once again, is forcing us to confront a modus vivendi that might have run its course. States and polities keep granting rights to people with no expectation that, except perhaps in some key circumstances, full reliance on them be thwarted by frictions, laziness, or lack of interest. This is the same story in relation to data protection and copyright:6 a legal framework that, until now, was never invoked at scale.
But if you change these parameters, if you put a wealthy and often unliked multinational in the role of the helpless defendant, give every aggrieved person a one-click cause of action, then this calculus won’t hold anymore. In that sense AI does not break the law, but it finally makes us apply the one we already had. We may find that this is more than we bargained for.
Hypergraphia Meets Autocomplete
Every profession that produces text (“manipulate symbols”) for a living is subject to the same temptation of using the AI shortcut anyway. This column has catalogued, of course, lawyers, but also academics and doctors falling into the traps of AI misuse.
The most recent firm to put its head in the stocks is KPMG, and on the most ironic subject possible; the FT reports:
A KPMG report on how AI is being used by businesses across the world exaggerated adoption of the technology with bogus case studies that appear to have been based on AI hallucinations. The October report, “Redefining excellence in the age of agentic AI”, made numerous false claims about the use of AI by organisations including the Swiss bank UBS, the UK’s National Health Service and the public transit groups Swiss Federal Railways and Transport for London.
The inaccuracies were identified as AI hallucinations by the research group GPTZero and verified by the FT.
This is not the first (and certainly not the last) such incident: EY got into hot water on the exact same grounds just last month, and before that, that was Deloitte made to reimburse clients over reports full of hallucinated material.
But one thing that struck me here is the delay: this report was available on KPMG’s website since October, sat there for months, and the hallucinations surfaced only now, because GPTZero (who also located the issues with EY’s earlier report) looked for it deliberately and blew the whistle.
This suggests that anyone who actually read that report earlier missed the hallucinations or did not care. But even that assumes that someone read that report, and this assumption may not hold. What is, after all, the value of such reports ? In conveying information you could find online or figure out by yourself ? And/or showcasing the skills and “au courant”-ness of the author ? None of this requires actual readers.
Another example: a few weeks ago, South Africa’s government put forward a draft AI policy that, again most ironically, was chock-full of hallucinations (I verified them myself). The draft was later pulled, but you can still find secondary material from reputable law firms advising their clients about it, as if nothing happened - epistemic pollution in action. But will the fake sources make a difference in terms of policy-making ? More or less than whatever true or false ideas and conceptions policy-makers already had in the back of their mind.
Not suggesting that it does not matter: obviously, don’t put fake sources in the wild. But everything is more or less grave, and while fake jurisprudence undermines a very important principle of legitimacy (i.e., that you can retrace a legal position to laws legitimately adopted or enforced by a legitimate enforcer), and fake legislative material might steer policy-makers in wrong directions, fake opinion pieces in a consulting memo are closer to the end “meh” end of the reaction spectrum. They are embarrassing, but did not necessarily move the needle.
And they can stay unnoticed forever, because no one cares about reading anyway.7
Every year some group of students during my coding course pick, as a final project, some kind of “let’s find which lawyers perform best in this particular jurisdiction/practice”, and it’s now my ritual to make them land on the important point that “winning” is never straightforward, and the best lawyers might actually take the hardest cases.
Cue to this idea that corporate profit is to a large extent based on the fact that people being bad at knowing what their money is actually buying or not, and that AI might actually increase information in this respect, to the advantage of the customers.
I am struck that Harvey said they were inspired by Cursor’s development of Composer: I never use Composer, and I am not sure I am ready to try and trust it for the kind of coding I do. This being said, and this comes back to my argument, I may be comfortable with it as some kind of under-the-hood tool that other, smarter models orchestrate.
The term “Large Libel Models” is borrowed from Eugene Volokh’s excellent paper on this subject.
And to that list can be added the “right to be forgotten” - good luck excising your past from a model’s weights !
But not you, dear reader, if you made it to the end of this blog post.