AI and The Ordinance of Reason

A Reply to Prof. Jeremy Christiansen

The New Digest is delighted to present a contribution by Kimo Gandall, Jack Kieffaber and Kenny McLaren, in reply to a previous essay by Jeremy Christiansen, “Congress.AI.” Messrs. Gandall and Kieffaber are recent graduates of Harvard Law School; Mr. McLaren is a computer scientist.

The natural lawyers and the judicial futurists were bound to clash eventually. And last week, it happened—with Professor Jeremy Christiansen firing the first shot in a terrific piece contemplating the validity of laws promulgated by “Congress.ai” (without consultation from any human actor). Given the number of em-dashes found in the Big Beautiful Bill, Professor Christiansen’s hypothetical might not have been one at all. And he didn’t think such laws would be valid—specifically on the ground that “law is an ordinance of reason, not the outputs of 1s and 0s.” It’s a convincing argument: Law must be reasoned to be valid, robots can’t reason, so law promulgated by AI is not valid.

Here’s another argument: “pens and papers can’t reason, therefore laws promulgated by pens and papers is not valid.” That one’s less convincing. But our job today is to convince you that these are the same argument. And that’s because an AI is simply a method of enumeration — the same as speaking or writing. The only difference is that, while oral and written traditions enumerate specific laws, an AI tradition would enumerate the very ordinances of reason that give rise to those specific laws; in short, it would enumerate a law about how to make laws. This nuance makes AI lawmaking more tethered to the natural law, not less. And classical lawyers should view this as both a blessing and a challenge; in teaching AI models how to write good laws, we have an unprecedented opportunity to memorialize our decisionmaking processes. And if we don’t do it, somebody else will — and we doubt he’s read Aquinas.

To that end, we’re going to show you one of our models. Like any AI, it's designed to perform a task; we’re going to demonstrate how we coded the model to perform that task, and how each spot of code represents a legislative decision as to how the task ought to be performed. The bot in question is a simple one, called “SCOTUS.ai.” Its task? Accurately predict Supreme Court decisions based strictly on the briefs. In teaching the bot to perform that task, we had to imbue the bot with value decisions about what makes a brief a winner or a loser — meaning we had to guess at the value judgments of the learned nine. Maybe we didn’t accurately reflect the court’s judgment (unlikely given that our model had 99% accuracy over two different terms); to the extent we did reflect said judgment, maybe that judgment is bad as a moral matter. What matters is that we enumerated such reasoned judgment as the input for—not the “product of”—1’s and 0’s.

And listen: We’re going to show you our code. It’s going to look scary, because it’s in a language you don’t speak. But even freshman compsci majors can read this stuff. It is a language — one far more precise than English. And it will be used in your lifetime — either by accident or on purpose — to enumerate the very value judgments that animate the natural law.

From the top: We had to teach the bot which inputs to receive — that is, we had to teach it which markings on a brief count as words. So, every merits brief our judge bot sees is forced through a deterministic pipeline—g(t) = norm ∘ token ∘ stem ∘ vocab(t)—that lower-cases, vectorises, and then filters the text through a comprehensive dictionary. Congress.ai would also have to take in some input—specifically, facts about the world, digestible via any combination of textual, video, and audio stimulus.

Once we taught the bot which words to absorb, we then had to teach it how to tell the winning words from the losing words—a process called “feature extraction,” where the bot decides what meaning to extract from the words. These are the metrics we deemed relevant to determining SCOTUS win likelihood (as seen above): analytic weight, authority, authenticity and emotional tone.

We thus decided to prioritize rhetorical hierarchy and emotional tenor alongside propositional content—which is a legislative decision about which briefs ought to win, namely those with tone and status. Again, one could easily argue that these are not the sorts of things the Supreme Court should consider. But that’s what we memorialized in the code — and it’s what we could amend if we wanted. Lastly, we had the decision of how complex to make the model; how much generalization error is society willing to tolerate in exchange for predictive power? If we increased the parameters by changing 1224 → 2226 (i.e. make the model bigger) my raw accuracy will increase; but on validation data (data the model has not seen), the generalization error may also increase. In legal terms, it is the same trade-off courts face between bright-line rules (predictable but rigid) and standards (flexible but fuzzy). In mirroring the Supreme Court, we tried to pick a middle ground — a decision both strict formalists and loose purposivists would deem equally and oppositely imprudent. But that’s the choice we made with this line of code:

Congress.ai, too, would have to decide what features to extract from the facts it receives about the world—defining the features that denote a good factual occurrence, features that denote a bad one, and determining how much good can be sacrificed to eradicate a particular bad. That might sound platitudinous, but policy experts are already in the business of extracting and weighting salient features from factual data. Look no further than the administrative state; for example, independent dispute resolvers for Medicare and Medicaid weigh nuanced metrics such as qualifying payment amount when determining the final insurance payment for a medical out-of-network service. Policy experts know the weights they’re looking for, or at least the ones that are in the conversation; Congress.ai would simply be an exercise in synthesizing and codifying agency-style expertise.

Maybe all of that was too technical. But we wanted to show you firsthand: Every single AI makes these kinds of decisions. And notice that the decisions are very granular; the task ahead is not just to enumerate moral platitudes, but to articulate them at the level of input recognition and output synthesis. The AI that drafted chunks of the Big Beautiful Bill was, in all likelihood, ChatGPT or Grok. These same types of decisions—what inputs the machine should receive, what factors it should screen for in converting those inputs to textual outputs, and what restrictions it should attach to those outputs—were baked into those models. And while they might strike you as more technical than moral, they determined what the law was; changing these vectors and weights would’ve resulted in different laws. And those vectors were set by humans.

The problem Prof. Christiansen is really pointing out, then, is one of locus. Somebody’s human reasoning dictated the phrasing in the Big Beautiful Bill—and that somebody almost certainly wasn’t a congressman. Rather, it was a thousand guys working at OpenAI over the course of a decade. That’s your new legislature. And it’s probably going to end up being your court, your agency, and your kid’s preschool teacher too. None of this means that the bots aren’t manifesting specific ordinances of reason. It just means that we didn’t vote for the reasoner—and, more fundamentally, that we might not like his reasoning.

So what’s a classical lawyer to do? We think it’s obvious: start seizing these new tools of enumeration. This moment is tailor-made for classical legal education; the laws of the future hinge on whether we can teach the smartest supercomputer that’s ever lived right from wrong in the natural sense. We endeavor to teach our students this; indeed, we endeavor to teach our children this. If we can’t teach it to an algorithmic superintelligence, skeptics will cast serious doubt on whether we’re capable of enumerating our beliefs at all—in any language. We can’t hide behind the mantra that “the products of 1’s and 0’s aren’t reasoning.” Human reasoning itself is the product of neuron and synapse, two doodads not unlike the 1 and the 0 in their rhetorical insignificance. And, as with AI, the products of our reasoning process are downstream of our inputs—our first principles, our priors. History has chosen us, in this unique moment, to write those priors down.

To close, we think it plain that laws produced by AI are still products of human reasoning, and thus still law in the Aquinian sense. But even if they’re not… there’s also that Hartian sense. The one that says a law is whatever they’ll kill you for not following. The Hartian sense isn’t very spiritually satisfying. But take this to the bank: If you try and defy Congress.ai’s new Big, Beautiful Bill on the grounds that it’s not an ordinance of reason, you’re going to jail. Like it or not, AI is going to play a heavy—and perhaps total—role in lawmaking within our lifetimes. We can cry about it. Or we can get ahead of it; right now we’re behind, and the fellas in the lead don’t read this substack. The classical lawyer must become a judicial futurist if the natural law is ever to practically govern again.