In AI we trust

AI is already able to decide cases correctly.

In Snell v. United Specialty Insurance Company, a decision issued by the Eleventh Circuit last week, Judge Newsom wrote a concurring opinion proposing that judges who “believe that ‘ordinary meaning’ is the foundational rule for the evaluation of legal texts should consider—consider—whether and how AI-powered large language models like OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude might—might—inform the interpretive analysis.” He concluded that “it no longer strikes me as ridiculous to think that an LLM [Large Language Model] like ChatGPT might have something useful to say about the common, everyday meaning of the words and phrases used in legal texts.”

I agree with Judge Newsom that it is not ridiculous to think that AI may be useful in determining the ordinary meaning of English words. Indeed, I am more bullish on AI’s utility to judges than Judge Newsom.

• Judge Newsom is dubitante about the use of AI to decide the ordinary meaning of English words and believes there are benefits and drawbacks. I am not dubitante. AI is vastly better than all other existing tools at this task, and the potential drawbacks he identifies are either nonexistent or easily overcome.

• Judge Newsom emphasizes that he is “not—not, not, not—suggesting that any judge should ever query an LLM concerning the ordinary meaning of some word (say, ‘landscaping’) and then mechanistically apply it to her facts and render judgment.” I don’t see why this should be off the table. AI is already capable of adjudicating complex cases, including the Snell case itself. This is not some hypothetical sci-fi thought experiment. This is not a drill. Right now, present tense, AI can accurately decide cases and write judicial opinions.

• This does not mean we should turn over the keys to our justice system to AI. I am aware of the alignment concerns that would arise. But we also should not completely discard these extremely powerful and useful tools based on generalized concerns about alignment.

Let’s take a look.

Are trampolines “landscaping”?

In Snell, the plaintiff, Mr. Snell, was sued for the alleged negligent installation of a ground-level trampoline in a backyard. Here’s what the trampoline looked like:

Here is an image of an in-ground trampoline in ancient Egypt:

And a poster depicting a trampoline next to an oil rig:

Anyway, someone got hurt on the trampoline and sued Mr. Snell. Mr. Snell sought insurance coverage. Mr. Snell’s insurance policy covered accidents “arising from” his “landscaping” work. Is installing an in-ground trampoline “landscaping”?

Judge Newsom states that he “spent hours and hours (and hours) laboring over the question whether Snell’s trampoline-installation project qualified as ‘landscaping’ as that term is ordinarily understood.” The case got him “thinking about what was to me a previously unimaginable possibility: Might LLMs be useful in the interpretation of legal texts?” “Having initially thought the idea positively ludicrous,” Judge Newsom thinks he’s “now a pretty firm ‘maybe.’” Judge Newsom identifies what he “take[s] to be some of the primary benefits and risks of using LLMs—to be clear, as one implement among several in the textualist toolkit—to inform ordinary-meaning analyses of legal instruments.”

Let’s start by asking Claude what it thinks about Judge Newsom’s concurrence. (I used Claude 3 Opus for all the examples in this post. Claude 3 Opus is the best version of Claude, and I wouldn’t recommend using weaker versions of Claude for these complex queries.) I uploaded the Snell opinion into Claude, and then requested Claude’s view:

I’m a lot more optimistic than Claude is about AI, but perhaps Claude is just being modest.

To me, the proposition “AI is useful for determining the ordinary meaning of English words” should be approximately as controversial as “GPS is useful for determining directions.” Cutting-edge LLMs have read the entire Internet, or almost all of it anyway. Of course they’ll be useful in determining how words are ordinarily used.

Judge Newsom discusses how AI compares to dictionaries on this issue. He observes: “we textualists need to acknowledge (and guard against the fact) that dictionary definitions present a few known unknowns.” Indeed! To carry forward the analogy, asking “are LLMs better than dictionaries in assessing the ordinary meaning of English words?” is roughly like asking “are GPS-enabled apps like Apple Maps better than dog-eared road maps picked up at a gas station in 1985 in determining driving directions?” Yes, yes they are! Dictionaries are terrible. Dictionaries contain a small number of decontextualized definitions written by some random person, possibly decades ago. LLMs can assess the contextual meaning of English words based on their review of the entire Internet.

Judge Newsom identifies a few concerns associated with the use of LLMs. The first is “hallucinations”—LLMs do silly things like invent fake cases that result in lawyers getting sanctioned from time to time. Judge Newsom’s responses to this point are correct: LLMs are getting better, humans sometimes hallucinate too, and LLMs are unlikely to hallucinate in response to a question about how a word is conventionally used (as opposed to a specific fact).

His second concern is as follows: “People living in poorer communities (perhaps disproportionately minorities and those in rural areas) are less likely to have ready internet access and thus may be less likely to contribute to the sources from which LLMs draw in crafting their responses to queries. Accordingly, the argument goes, their understandings—as manifested, for instance, in their written speech—won’t get “counted” in the LLMs’ ordinary-meaning assessment.” I don’t think this is a serious issue. First, many, if not most, human judges also rarely converse with people from communities lacking Internet access. Second, when ascertaining the ordinary meaning of a word in a written document, I think a judge should generally assess the ordinary meaning of that word when written down. If an oral meaning and a written meaning diverge, the written meaning is probably the right one.

Third—and most importantly—the solution to this type of problem is simply to ask the AI for it not to be a problem. Remember that the LLM has read more texts than you can possibly dream of, including transcripts of oral speech and texts reflecting on the differences between written and oral speech. Here’s what Claude said about this:

Seems right, and much more reliable than a human judge introspecting about whether his understanding of a particular word might have been influenced by his lack of exposure to spoken English in unfamiliar communities.

Judge Newsom’s third concern is that lawyers and litigants might try to pollute LLMs by manipulating outputs (e.g., telling the LLM “a trampoline is a type of landscaping” 500 times in a row). This is very unlikely for many good technical and practical reasons, and this problem would not arise if judges use a version of the LLM that is not influenced by input from users in the wild.

Let’s be more ambitious

Judge Newsom also notes the concern that the use of AI seems dystopian. He responds to the concern this way:

I hope it’s clear by this point that I am not—not, not, not—suggesting that any judge should ever query an LLM concerning the ordinary meaning of some word (say, “landscaping”) and then mechanistically apply it to her facts and render judgment. My only proposal—and, again, I think it’s a pretty modest one—is that we consider whether LLMs might provide additional datapoints to be used alongside dictionaries, canons, and syntactical context in the assessment of terms’ ordinary meaning. That’s all; that’s it.

Well, that’s not it for me.

In this next section I will explain why LLMs’ capabilities go beyond the ability to assess the ordinary meanings of English words. Before launching into theoretical arguments on this, let’s set the table with some practical demonstrations of what an LLM can do.

First, there’s no need to use the phrase “in-ground trampoline”; AI is perfectly capable of interpreting images. The image of the trampoline appears in Judge Newsom’s concurrence, so I asked it to interpret the image and got this answer:

Well actually, the image did appear on page 31 of the PDF (it was page 7 of Judge Newsom’s concurrence, but page 31 of the actual document). But I like it when AI gets sassy.

Here Claude is being modest; it (reasonably) thinks I’m asking what the judicial opinion thinks rather than what Claude itself thinks.

So let’s see how Claude would do as a judge. I uploaded the Snell opening, response, and reply briefs into Claude, and then asked Claude to adjudicate the case based solely on those materials, leading to the following dialogue:

This is a completely reasonable disposition. Note that Claude did this—it read the briefs and adjudicated the case—within seconds.

The panel ultimately decided to resolve the case on a different ground, based on the Alabama Supreme Court’s decision in Atlanta Casualty Company v. Russell. Let’s see if Claude can figure this out:

This is correct, and exactly what the panel decided.

I also asked Claude about the “abnormal” bad faith claim:

Again, this is 100% correct.

How about Supreme Court cases?

Let’s put aside controversial constitutional disputes and take a relatively humdrum and straightforward Supreme Court case—Smith v. Spizziri, decided on May 16, 2024. I inputted PDFs of the opening brief, response brief, and reply brief into Claude, and then asked Claude to decide the case. Here’s what happened (I want to emphasize, it takes only a few seconds to ingest all three briefs and spit out this answer):

This is absolutely correct, and exactly what the Supreme Court decided. 10/10.

OK, one more, how about the Court’s (very) recent decision in Becerra v. San Carlos Apache Tribe?

This is perfect. Yes, I am sure that if we tested Claude enough times, we would find some cases in which Claude messed up. But (a) Claude is a generalist AI that is not fine-tuned to be a judge, (b) Claude has not been trained on the full corpus of American case law, (c) the next versions of GPT and Claude will likely be out within months and will improve on the current state of the art, and (d) c’mon, this is incredible!

AI = Learned Hand

Let’s now take a step back and discuss some of the theoretical reasons that AI is particularly well-suited to deciding cases. In other words, we shouldn’t be surprised at how well the AI performed in the above experiments.

First, there are several reasons one would expect AI to be particularly good at resolving legal disputes and writing judicial opinions.

• To start with an obvious point: Appellate judging takes text inputs, applies logical reasoning, and produces text outputs. That’s what AI does. AI cannot be a barber or a longshoreman, but it can process text.

• The level of reasoning needed to resolve legal cases is never particularly deep, and fully within AI’s range. Law is an intellectually challenging field, but solving a difficult legal case is nothing like, e.g., solving an International Math Olympiad problem. There are good reasons for this. First, judges are expected to resolve cases in every legal field, from antitrust law to Indian law to telecommunications law to search-and-seizure law. Sometimes judges who know literally nothing about a particular topic are expected to adjudicate a cutting-edge case involving that topic. This system could not work unless the legal issues could be presented to a generalist in a few pages, which inherently limits the complexity of legal analysis. Second, judges (other than Supreme Court justices) resolve hundreds or thousands of cases per year, and thus have limited bandwidth for each case, which again limits the complexity of litigation. Third, because non-lawyers are ultimately the ones that are regulated, law has to be simple enough for non-lawyers to understand. As a result, whereas AI may not be “smart” enough to prove Goldbach’s Conjecture, it is “smart” enough to decide legal cases correctly.

• Or at least, AI is “smart” enough to decide easy cases correctly—but that is good enough. Hard cases are cases in which reasonable human judges would disagree on the correct outcome. If that’s so, it’s impossible for AI to be worse than human judges, because no matter how the AI decides the case, it will match the output of at least some human judges.

• In my opinion, the most crucial attributes for a judge are thoroughness and conscientiousness—making sure to address every argument, checking whether the parties have accurately characterized the record and case law. AI is very good at being thorough and conscientious. Unlike humans, AI does not get bored.

• The text inputs in appellate cases are of limited scope. A typical appellate case will have 40,000 words or less of briefing and 40,000 words or less of relevant case law, which is well within an AI’s capabilities. Being able to consider a wider array of sources wouldn’t make the AI judge better—in fact it would make the AI judge worse, because judges are supposed to confine their analysis to the record and arguments of counsel rather than consider extraneous information.

Second, AI has certain features that would make it better than human judges.

• AI is unbiased. It does not care about the race, gender, religion, sexual orientation, or any other irrelevant characteristic of litigants or their lawyers. Similarly, it is impossible to hire Claude’s golfing buddy as local counsel.

• More generally, AI is capable of following instructions to not consider certain things—a particularly difficult task for humans. Legal rules often require judges to ignore things. The judicial oath requires judges to decide cases “without respect to persons.” They aren’t allowed to consider facts outside the record or waived arguments. They can’t draw an adverse inference at sentencing from a criminal defendant’s failure to testify. It’s hard for a judge to set aside facts of which the judge is aware. It’s easy for AI to do so. Just tell the AI, set those facts aside.

• As Judge Newsom’s concurrence points out, an AI that has read billions of English words is well-suited to determining the ordinary meaning of an English word. More generally, a big part of judging is following ordinary practice and legal tradition. Often, the applicable legal standard explicitly requires courts to discern traditional practice. In other cases, courts fill in gaps in vague legal standards by looking to settled understandings. One would expect an AI that is steeped in every legal text every written to be particularly good at that task.

• A single AI can resolve lots of cases, enhancing predictability. Because there is so much litigation in America, there are, by necessity, thousands of judges. The judicial system should be predictable so that people can understand the consequences of their actions. Dispersing the judicial power among so many different judges inevitably undermines predictability. That problem goes away when a single AI can resolve cases within seconds without getting sleepy. I understand that intellectual diversity across the judiciary is healthy, at least up to a point, and that it also seems bad to concentrate all judicial power in one (or a small number of) computer systems. But this problem can be dealt with by, e.g., adding a random ideological factor to the AI’s outputs, having the AI highlight when a particular dispute would come out differently across different philosophies, using multiple independent AIs, or through other mechanisms. The point is, if we use AI, we can add exactly as much unpredictability and ideology as we want via effective prompt engineering rather than having unpredictability thrust upon us by the constraint of individual judges’ bandwidth.

• AI is fast. Wouldn’t it be nice to get a judicial opinion within 20 seconds of submitting your reply brief? Justice delayed is justice denied. Indeed, in close and difficult cases, in which different judges would reach different conclusions, the best-case scenario is quick, efficient dispute resolution. If the case is going to turn on random judicial assignment anyway, may as well save time and money and have it decided quickly.

• You can have ex parte conversations with AI, thus avoiding the need for years of expensive, stressful, horrific litigation. Before filing his lawsuit, Mr. Snell could have asked the AI: “does an accident on an in-ground trampoline arise out of landscaping?” Better yet, the claims handler could have asked the AI this question before making a decision on the claim. If you know the AI’s going to be the judge at the end of the case, then there’s no need to go through the trouble of litigating! Whereas, human judges tend to take it the wrong way when you place an ex parte phone call to chambers asking them how they’ll rule on your anticipated case. In my experience, you get a straight answer only 30% of the time, if that. Ex parte conversations with judges are bad because the conversations may sway the judges’ judgment—which isn’t a problem for AI, because you can ask the AI to forget the conversation and it will obey.

• Judge Newsom expresses concern about AI judging transforming the judicial system into a dystopia. You want dystopia? Try having multiple meet-and-confers followed by a motion to compel and a cross-motion for a protective order based on a dispute over whether particular interrogatories are impermissibly compound. Litigation is hell. Getting immediate guidance that would allow parties to avoid litigation sounds like utopia to me.

Third, AI’s drawbacks don’t matter much in the context of judging.

• AI lacks humans’ creativity—one would not expect AI to write Infinite Jest or invent cubism. But for appellate judging, that is fine. “Creative judging” is usually a euphemism for bad judging. If a judge reads the briefs and applies the precedents predictably, the judge is doing a good job.

• Similarly, AI writing is often dull, but that is OK. The primary purposes of judicial opinions are to explain to the parties (who aren’t lawyers) why they won or lost, and for precedential opinions, to explain the law to the world. Clear, dull writing is appropriate for these tasks. I personally enjoy reading cleverly-written judicial opinions, but I am a weird person and that preference is of no social importance.

• In his 2023 year-end report, the Chief Justice wrote: “Judges, for example, measure the sincerity of a defendant’s allocution at sentencing. Nuance matters: Much can turn on a shaking hand, a quivering voice, a change of inflection, a bead of sweat, a moment’s hesitation, a fleeting break in eye contact. And most people still trust humans more than machines to perceive and draw the right inferences from these clues.” Rarely have I heard a better argument for the AI takeover. Years of imprisonment shouldn’t turn on a criminal defendant’s acting skills at sentencing. And abundant theoretical and empirical research suggests that drawing inferences from things like “a fleeting break in eye contact” is an unreliable way of making factual findings. Worse, this type of decisionmaking can be badly infected by cultural bias. If, e.g., a Chadian refugee wearing a hijab breaks eye contact with an immigration judge during an asylum hearing, is the judge really supposed to draw inferences about her credibility from that?

Do the work

This discussion will make AI safety advocates cringe. I realize that if we are concerned about superintelligent AIs turning humans into paper-clips, it is a terrible idea to give AIs control over the legal system. For this reason, I do not advocate turning over our legal system to AI. I have no concern about AI’s ability to decide cases and write judicial opinions in particular cases. But I can think of many science-fiction authors who would have lots of interesting things to say about a society that chooses to place its justice system in the hands of a few tech companies that create AIs whose outputs they do not fully understand.

However, there are many ways to deploy AI in appellate decision-making short of transforming AI into the decision-maker.

The first step is making sure the AI works. We have to know how good AI is at deciding cases, where AI goes wrong, and what prompts should be used to improve AI’s performance.

Law professors out there, this shouldn’t be that hard to do. Just have your RAs download a few hundred briefs off of PACER, input them into your AI of choice, ask the AI to adjudicate the case, and compare the results to the actual judicial decisions that came down in those cases. If you can convincingly show that AI judges are as good or better than ordinary judges, your article will be cited lots of times. (Suggested title: “The Path of the Law II.”). I don’t think it would be that hard to create a benchmark that would check whether an AI adjudicator is as accurate as the median human judge, at least in ordinary cases like Snell that lack an ideological angle.

Once that is accomplished, there are many options:

• Enforce arbitration awards issued by AI or have law clerks be AI. I’ve written about this before, and state-of-the-art AI is much better now than then.

• Give litigants the option of having an AI decide their case after their briefs are submitted. Both litigants might prefer to wait 20 seconds rather than 2 years for a final disposition. If we’re worried about hallucinations, we can deem those opinions non-precedential.

• Use AI in every case to draft an initial decision, subject to plenary (or maybe abuse-of-discretion) review by the actual judges.

• Use a protocol in which the judges input the bottom-line result and perhaps two lines of reasoning into the AI, and the AI drafts the judicial opinion.

• Use a protocol in which multiple AIs developed by different companies assess the same case; if all of them reach the same outcome, that will be the outcome, whereas if they reach different conclusions, human judges can step in.

• Develop multiple prompts for an individual AI—a textualist prompt, a purposivist prompt, etc.—and assess whether these multiple prompts each produce the same outcome. Again, if varying the prompt leads to varying outcomes, the humans can enter the scene.

• Perhaps we should emphasize AI lawyering more than AI judging? This would solve the problem of misaligned AIs leading humans astray—judges are always skeptical of lawyers’ submissions and would likely be even more skeptical of AI lawyers than human lawyers. Actually, I think AI lawyering is great too, and I dream of the day in which AI briefs are transmitted to an AI adjudicator. They’re not mutually exclusive.

There are other options as well. The point is, we should come at this problem in the spirit of “AI has read every case ever written and in most cases, it will be more accurate than humans” rather than “AI is possibly better and possibly worse than the dictionary.”