Helpful bots are not helping
What makes a legal argument bad ? As in, something a better lawyer, or process, would immediately strike ?
There are several dimensions to that question. To start with, a bad argument might be something that’s just wrong on the law, certainly. But even that is not always obvious, as “wrong” admits a lot of degrees, while some “wrong” takes might be worth pushing for strategic reasons. Beyond this, a bad legal argument might be something that focuses on a tree and misses the forest; or a point that might backfire at a later juncture; or something that, when taken to its fullest logical conclusion, would eventually become detrimental to your position. Considerations of good faith, of diligence, and of being respectful of the other parties may all come into play as well to define, in given circumstances, a “bad” legal position.
We may also take the problem from the other end, and wonder what makes a good legal argument. But as discussed earlier about getting to a “good enough” stage, this question accepts no easy answer either: knowing how to stop improving - or even knowing whether one is improving - a legal brief is a skill that defies strict definition.
Still, this suggests a resolution, of sorts, to the question about “bad” legal arguments: knowing what makes legal reasoning bad is hard, but it might be easier to identify what can make it worse.
And in this respect, obeying the house tradition of always trying to compare law to code, consider Zach Kehs’s recent article pointedly entitled: “There’s No Limit to How Bad Code Can Get”:
Software is in the domain of the abstract. It is not like a building, or a bridge, that is in the physical realm where you can see and feel the nature of the thing. If you continue to add floors and rooms to a building forever, it will collapse. Software faces no such constraint. The code can always get worse. There can always be a new layer of indirection or a reduction in performance.
[…]
Metaphors like a “collapsing building” or a “sinking ship” are not appropriate for software, yet we can embrace them anyway to emphasize what makes software different. The building is infinitely collapsing. The ship is infinitely sinking.
What else is in the domain of the abstract ? Law.
And it shows at all layers. Legal argument, of course: although that is often bound by time (you file or send your memo by a deadline), there is some amount of repeatability involved, through templates, repeat cases, or even the legal works that span years or decades. You can (and we frequently do) make it worse as time stretches. Legislation partakes of the same spirit: amendments pile on, nothing is ever refactored, and the only way to reform a system is to build one in parallel.
Now, rather unsurprisingly, the ways AI vibe-coding and vibe-lawyering worsen your output are largely the same. They do so chiefly by adding things. In coding, this takes the shape of more abstractions, of infrastructure that is sometimes oversized for what you want to do, of a battery of tests that serve no purpose.
In law, this often means longer submissions and, in particular, constant, relentless hedging. Consider this recent order by a judge from the U.S. District Court for the District of Maine, reacting to an 80-page submission by a pro se defendant:
The proposed filings are clearly not the work of an organic mind with linguistic talents developed in the crucible of human experience. It is obvious that the entity that drafted them has an artificial intellect that is unconcerned with, for instance, getting to the point. To comment on the filings at length would only serve to reward them. Still, I offer some particularly glaring artificial ticks [sic] that are a complete waste of the reader’s time.
Throughout the motion, the author has taken special care to tell the Court what it is NOT arguing or suggesting. For instance:
“Defendant does not ask the Court to accept any of that . . ., and does not ask it to revisit . . . .”
“Defendant does not read that order as . . . .”
“Defendant does not ask the Court to decide . . . .”
“Defendant does not press a claim that . . . .”
“Defendant does not ask the Court to . . . .”
“Defendant does not offer that as an excuse for anything, and it is not one . . . .”
“Defendant does not assume what the Court would have done . . . .”
“Defendant does not contend that . . . .”
“Defendant does not ask the Court to . . . .”
“Defendant does not characterize . . . .”
These ten excerpts are drawn from the first 26 pages alone. Many, many more such empty statements follow.
This is indeed a rather recent AI tell, more subtle - but often more definite - than the “it’s not X, it’s Y” figure de style. LLMs, especially recent ones, constantly hedge and seek to ensure that their point is never taken beyond its limits, in contexts where a human would just assume no such hedging is needed. I have seen it in various contexts, be it in data analysis (being punctilious in laying out all the ways the data was unclean), or in legal memos (engaging in digressions about why a precedent they invoked was not a perfect fit and should absolutely not be taken too literally, god forbid). These concessions are useless when they are not detrimental.
Models may do so in the spirit of wanting to do good and follow good practices, but in the process forget that moderation is the mother of all virtues. The universe of what one is not asking is limitless; there is always a further hedge, and these 80 pages could have been 400 listing more of those. With the result that AI-generated legal arguments are not necessarily bad, but they can always be worse.
Caring bots needed
Followers of this newsletter for the past few months may have intuited that a few things, or themes, animate its author, such as:
The growing hypergraphia of our times, the mountains of text that cast their tutelary shadows over our lives;
The fictions we tell ourselves to make a system work, such as the idea that everyone actually reads and signs these mountains of text; or
The role of AI (and automation more generally) in shedding light on, and often belying, such fictions, forcing us to reckon with the issue.
This is an incomplete list, but hopefully it explains why I particularly appreciated this piece, “My Bot Can Sue Your Bot”,1 which observes that:
It is literally impossible for ordinary people to read, understand, compare, remember and manage the online contracts that govern their digital lives.
And yet courts routinely enforce them.
The result is a strange legal architecture built on fictional consent. […]Then along comes AI. What you can’t (or won’t) read, you can have your AI bot read for you. AI can read, compare, monitor, flag, decide, notify and sometimes transact. That changes the problem.
If companies can use bots to draft, update, personalize, present and enforce contracts, why can’t consumers use bots to read them?
And continues with the argument that, instead of agreeing to all these online contracts, people could use their agents to seek to read, flag, opt out and, when warranted, negotiate with (or sue) the services and persons on the other side of the fine print you never look at.
Which is a bit too cute, and glides too much over what makes this particular fiction alive. Which is not, as argued in the piece, that people are lazy, or lack time to care about their digital “rights”; but instead, as the scare quotes should indicate, that there is not much of a there, there. Like much else in this area, deploying agents for online contracts will be easy prey for the privacy paradox: you would still need to state what exactly you care about, and chances are you don’t know, or at least not to the point of being aware of the trade-offs this implies.
Still, it is a fact that there is a minority of people who care about this, and reducing inconvenience and frictions for these people may make action on their part easier. But then, this would raise questions about the robustness (and relevance) of the regime. Matt Levine made the financial version of this point recently: credit card rewards and US mortgages are generous largely because the people holding them won’t use them optimally. Comes an AI that refinances everyone’s mortgage the moment rates drop, and in all likelihood mortgages would be made more expensive for everyone.
In the legal realm, I have dubbed this the gym membership model of the law: it works as long as you don’t exercise your rights. When you do, the rules are bound to change. Which is why I am rooting for using bots to do the caring for us: that may force a reappraisal.
Nothing further, your Honour
When it comes to legal judgment, the naive view frequently overestimates the importance of getting it right. As in, the general public will bemoan the judge or arbitrator who, seemingly, made what we see as a mistake, or decry the process that led to what strikes them as the wrong outcome.
Instead, it is often as important - if not more - that we get an answer at all. This is the standard view behind the classical Latin maxim interest reipublicae ut sit finis litium, the notion of res judicata, or the more modern Jackson quip that, roughly, the Supreme Court is final not because it is right, but right because it is final. The law needs an answer more than it needs the perpetually latest answer.
And this tension between truth and finality runs through the entire legal process. Parties are typically ordered to brief and file their arguments by a given point or deadline, and strict rules then govern whether they can change these arguments or add to them. This is a notable constraint on correctness: maybe you needed more time, maybe you should have waited for new developments, etc. But those rules are there so that everyone focuses on one canonical version of the argument, for pragmatic reasons (what do you cite to ?), but also because due process requires it (you want no surprises).
Anyhow, that equilibrium works in law because, again, truth and correctness are not that important. But the relation also goes the other way round: truth and correctness yield to the pragmatic reality that we can’t keep track of an ever-changing document.
What if that was not true, though ? For instance, what if AI could be used to automatically and autonomously update a particular document to reflect changing circumstances ?
This is the idea behind recent suggestions to rethink how we write academic and scientific papers: not as a snapshot in time, like judgments or pleadings, but as something more dynamic, that adapts itself through time. Come new methods, or new data, and the paper you were reading yesterday may reach a new conclusion.
This is exciting, but also means research will be confronting a choice between the canonical text and the useful living derivative. That choice, legal systems made it centuries ago in favor of the snapshot, because the snapshot is what you can enforce, appeal, and execute against (the “useful living derivative”, here, would be legal commentary or doctrine, whose authoritativeness suffers precisely from being derivative). Nobody garnishes a bank account on the strength of a document that might have updated overnight.
And in line with these ideas, a team from Stanford recently published in Nature under the title “Reimagining research papers as interactive and reliable AI agents”. Extract:
Paper2Agent transforms research output from passive artefacts into active systems that accelerate use and discovery. Conventional research papers require readers to understand and adapt the paper’s code, data and methods to their work, creating barriers to dissemination and reuse. […] By turning static papers into interactive AI agents, Paper2Agent introduces a paradigm for knowledge dissemination and a collaborative ecosystem of AI co-scientists.
What would this look like in the legal domain ? As recounted by Osborne Clarke, a recent ICC panel went over related ideas, and in particular that of moving away from repeated (and repetitive) briefs towards the creation of a shared evidential record that could be queried by AIs - on both sides, but also by the tribunal. The parties’ arguments would then orient themselves around this shared context. And a new legal function would be located in deciding what, and when, gets frozen, and what can remain dynamic.
Whether that’s feasible or not, it illustrates an important point: that AI-led efficiencies do not necessarily stem from doing more of the same faster. Instead, they may allow entirely new approaches to the legal process, if we are willing to try them.
And very obviously AI-generated, but we here at Artificial Authority are feeling liberal-minded.