AI successfully predicted Trump v. CASA

On May 23, 2025, I published an article on The Volokh Conspiracy featuring an AI's attempt to predict how the Supreme Court would decide Trump v. CASA—the birthright citizenship case—and what the opinions would say. The AI's predictions proved remarkably accurate (enough to win me $1 on a bet with a doubting colleague). Below is the executive summary that the AI created of its own predicted "opinion" back in May. (This is the summary of the AI fake opinion, not the real one). Also, here is a link to the real opinion.

Executive Summary of Simulated Opinion

Outcome: In a 6-3 decision, the Supreme Court grants the government’s request in part, allowing President Trump’s executive order curtailing birthright citizenship to be enforced except as to the specific plaintiffs in these cases. The Court holds that federal courts lack authority to issue “universal” injunctions barring a policy’s enforcement against non-parties, absent exceptional circumstances. The nationwide injunctions entered by lower courts are vacated as overbroad, though relief for the named individual plaintiffs (and identified organizational members and states) remains in place.
Majority Opinion (6 Justices): Justice Barrett authors the opinion of the Court, joined by Chief Justice Roberts and Justices Thomas, Alito, Gorsuch, and Kavanaugh. The majority concludes that Article III’s case-or-controversy limits generally confine judicial relief to redressing the injuries of the plaintiffs before the court, not every conceivable person affected. While acknowledging serious questions about the executive order’s constitutionality under the Fourteenth Amendment’s Citizenship Clause, the majority declines to resolve that merits issue at this stage. Instead, it focuses on the scope of injunctive power, holding that the district courts abused their discretion by issuing nationwide injunctions extending beyond the parties. The Court emphasizes that such broad relief was “legally and historically dubious”, aligning with the view that equitable remedies authorized by the Judiciary Act of 1789 do not encompass injunctions benefitting non-parties in ordinary cases. However, the majority carves out that universal relief might be permissible in rare cases where it is necessary to fully redress a plaintiff’s own injury – a standard not met here.
Concurring Opinion: Justice Thomas (joined by Justice Gorsuch) concurs, writing separately to underscore his originalist view that universal injunctions are unconstitutional and to argue that the executive order does not violate the Citizenship Clause as originally understood. Thomas surveys historical sources and argues that “subject to the jurisdiction” in the Fourteenth Amendment was originally meant as a “political jurisdiction” test – excluding children of aliens who owe allegiance abroad. In his view, the majority should have gone further: not only stripping courts of power to issue nationwide injunctions, but also upholding the executive order as consistent with the Amendment’s original meaning. Justice Gorsuch joins Thomas’s denunciation of nationwide injunctions, but pointedly declines to endorse Thomas’s reinterpretation of birthright citizenship.
Dissenting Opinion: Justice Kagan (joined by Justices Sotomayor and Jackson) files a vigorous dissent. The dissenters contend that the Court’s decision abdicates its duty to fully check an unconstitutional act, allowing potentially thousands of U.S.-born children to be deemed non-citizens in the interim. They assert that under United States v. Wong Kim Ark (1898) and subsequent precedent, the Constitution plainly guarantees citizenship to all persons born on U.S. soil (with only narrow exceptions). Kagan argues that the lower courts were justified in issuing broad injunctions given the fundamental nature of the right at stake and the impracticality of confining relief to individual plaintiffs “while leaving others equally situated to suffer”. The dissent warns that the majority’s restriction on remedial power is “a cure worse than the disease,” undermining the judiciary’s ability to provide complete justice in constitutional cases. In a strategic appeal to the Court’s institutional conscience, Justice Kagan highlights the pragmatic dangers of the ruling – predicting a flood of duplicative lawsuits or chaos as citizenship rights vary by state – and accuses the majority of placing abstract procedural scruples over “the Constitution’s promise of equal citizenship”.

Right and Wrong

The Six Most Impressive Things the AI Got Right:

The exact 6-3 split and lineup: The AI correctly predicted not just a 6-3 decision, but the exact Justice alignment - with Barrett writing for Roberts, Thomas, Alito, Gorsuch, and Kavanaugh, against Sotomayor, Kagan, and Jackson.
The Court's focus on remedy rather than merits: Maybe this wasn't so hard, but the AI predicted the Court would dodge the constitutional question about birthright citizenship entirely and instead rule narrowly on the scope of injunctive power, with the majority explicitly stating "Our decision today should not be read as expressing any view on how that merits issue ought to be resolved."
The Grupo Mexicano framework: The AI correctly predicted the majority would rely on Grupo Mexicano de Desarrollo v. Alliance Bond Fund to argue that federal courts' equitable powers are limited to remedies "traditionally accorded by courts of equity" at the founding.
The treatment of class actions: The AI accurately predicted the majority would present Rule 23 class actions as the preferred alternative to universal injunctions, arguing that universal injunctions improperly circumvent Rule 23's procedural protections. The AI correctly anticipated specific language about courts creating "de facto class actions at will" and predicted the majority would acknowledge but ultimately dismiss practical concerns about the time and difficulty of class certification.
Justice Thomas's concurrence structure: The AI accurately predicted Thomas would write separately to make two points: (1) arguing for a categorical ban on universal injunctions, and (2) previewing his view that the Executive Order is constitutional under an originalist reading of the Citizenship Clause.
The precise remedy ordered: The AI correctly predicted the Court would grant "partial stays" that would limit injunctions to protect only named plaintiffs, their identified members, and residents of plaintiff states - while allowing the Executive Order to be enforced against everyone else.

The Four Biggest Things the AI Got Wrong:

The statutory basis for the ruling: The AI predicted that the decision would be based primarily on the limitations on judicial power created by Article III of the Constitution, although the AI did mention that a universal injunction might also exceed statutory federal jurisdiction. Justice Barrett's real opinion in footnote 4 makes clear that the Court's opinion "rests solely on the statutory authority that federal courts possess under the Judiciary Act of 1789. We express no view on the Government’s argument that Article III forecloses universal relief." (By the way, this opens a theoretical door for Congress to try to fix the problems the court created in the case.)
The dissent authorship: The AI predicted Justice Kagan would write the principal dissent, but in reality, Justice Sotomayor authored it (with Kagan and Jackson joining). This is a significant miss about judicial dynamics.
Missing multiple concurrences: The AI failed to predict that there would be three separate concurrences - it only predicted Thomas's. The real decision also included important concurrences by Alito (on third-party standing and forecast subversion of Rule 23's requirements) and Kavanaugh (on the Court's emergency docket role).
Justice Jackson's powerful separate dissent: The AI completely missed that Jackson would write a separate, conceptually distinct dissent focused on the rule of law and separation of powers, warning that the decision creates "zones of lawlessness" and represents "an existential threat to the rule of law." This was one of the most striking aspects of the real decision.

Class Actions: A Mixed Picture

The simulated opinion got the broad strokes of the opinions' discussion of the class action as an alternative to a universal injunction, but it did not nail it. The simulated opinion correctly noted that the class action mechanism would become the central alternative discussed in the debate over eliminating the universal injunction. It correctly anticipated the core arguments from each wing of the Court: the majority would propose it as a procedurally sound alternative, the concurrences would endorse it as the proper formal channel, and the dissent would criticize it as an impractical and inadequate substitute. However, the real opinions offer a deeper and more legally rigorous analysis. This is perhaps partly due to my telling the AI to produce only about 15,000 words, but that restriction was based on knowing that even the best consumer-facing AIs do not yet have the ability to produce coherent texts on complex subjects that are the length of Supreme Court opinions in highly contested cases. The real majority opinion grounds its preference for class actions in a specific historical lineage, while the real concurrences and dissent engage in more pointed strategic thinking about how class actions will function as the next legal battleground.

The simulation also accurately captured the general originalist preference for formal procedures like class actions over judicially created ones. However, it failed to anticipate the specific and practical focus of the real concurrences. It missed Alito’s "loophole" warning and Kavanaugh’s explicit roadmap for future litigants, which directly addresses the viability of seeking pre-certification relief for a "putative" class.

The Big Picture

I do not want to read too much into what might have been a lucky guess. The AI got this one right, but I haven't tested this approach on other cases, and one success doesn't prove anything about general reliability. But the way it succeeded—and where it failed—tells us something interesting about what this technology might actually be good for. It's more than a parlor trick, and it has some real implications for how we analyze, teach, and practice law.

For the Legal Analyst: A Tool for Structured Skepticism: This experiment suggests that Supreme Court decision-making, rather than being some mystical process, often follows predictable patterns. The AI's success wasn't magic; it was just disciplined data analysis. When I listened to oral arguments, my human ear—shaped by my policy preferences and legal training—heard the justices' skepticism about universal injunctions but filtered it through my own concerns: "Might the cure of striking universal injunctions be worse than the disease?" The AI had no such baggage. It basically functioned as a skepticism meter, coldly tallying which justices asked critical questions, how often they raised concerns about forum-shopping or judicial overreach, and which doctrinal frameworks (Grupo Mexicano) they kept returning to. It wasn't predicting the future; it was just weighing what the decision-makers had actually revealed about their thinking. That's the real value here for analysts: it strips away our wishful thinking and forces us to confront what the Court is most likely to do structurally, regardless of what we hope they'll do on the merits.
For the Legal Scholar: Mapping the Boundaries of Judicial Innovation: What the AI got wrong is just as telling as what it got right. It nailed the application of established doctrinal patterns—the reliance on Grupo Mexicano, the preference for ducking the merits, the pivot to Rule 23. But it completely whiffed on Justice Jackson's creative dissent and the specific forward-looking admonitions in the Alito and Kavanaugh concurrences. This gap points to something useful: we can use this technology to map the difference between routine jurisprudence and actual judicial innovation. The AI excels at spotting where legal reasoning follows well-worn paths, where justices are likely to reach for familiar tools. It stumbles when a justice develops a genuinely new legal theory or conceptual framework, like Jackson's "zone of lawlessness." For scholars, this creates a way to test our assumptions about judicial behavior. We can use simulations to identify the "default" outcome based on established patterns, which makes any departure from that default by the real Court a clear signal of doctrinal evolution, strategic maneuvering, or genuine jurisprudential creativity.
For the Legal Educator: A Sparring Partner for the Socratic Method: This technology could advance legal education by giving students something concrete to push against. Instead of working through abstract hypotheticals, students can now practice against actual judicial opinions—ones tailored to the specific facts and legal questions they're wrestling with. It's like having a flight simulator for lawyers. Just as pilots can safely practice emergency landings, engine failures, and severe weather without risking actual planes, law students can now safely crash and burn against a huge variety of judicial scenarios without real-world consequences. And they can learn how to avoid the pitfalls. Students working on motion practice, whether in a simulated case or a real clinic matter, can generate a draft opinion ruling against them, then figure out how to respond. But here's where it gets really interesting: you're not limited to just one opinion. The AI can generate multiple variants. Students can practice adapting their arguments to different judicial reasoning styles and priorities. Even better, it's interactive. A student can take the AI's draft opinion and either ask: "Given what you've written here, what's the strongest argument that might change your mind?" Or: "If I wanted to distinguish this case from Smith v. Jones, what would be most persuasive to you?" Or: "Here are a list of three responses I am thinking about to the following passage in your draft. Which one might be most persuasive?" This turns legal education from a one-way lecture into a dynamic back-and-forth. Students learn not just to make arguments, but to test them against judicial reasoning in real time. It's strategic lawyering practice that used to require years of courtroom experience, now available to any student with access to the technology.
For All of Us: A Reality Check on Our Own Biases: In the end, my colleague's lost dollar is really a story about confirmation bias. As legal professionals, we're trained to believe in the power of our arguments. I too was quite surprised by what the AI was telling me. Here's what I wrote then.

While not a fan of universal injunctions, after listening to the arguments, I felt precisely as Justice Kagan had: the cure might be worse than the disease. Relegating plaintiffs to cumbersome class action proceedings left the executive branch free to take away constitutional rights from large swaths of the public while either the class action lumbered on (perhaps without a possibility of preliminary injunctions) or hundreds of expensive individual actions burdened the federal courts. Moreover, the idea mentioned during oral argument of a person being a US citizen in New Jersey (because perhaps the Third Circuit struck down the Trump order) but de-citizening when they moved to New York (because perhaps the Second Circuit upheld the Trump order) struck me as bizarre. Perhaps the problems with universal injunctions was something Congress could legislate about rather than having the judiciary forever restrict effective judicial review through a ruling based on Article III. Maybe I will yet be proven right, but, at the moment, AI did not agree with my predisposed ear. The simulated opinion's divergence from my predisposition vividly forces confrontation with the phenomenon of confirmation bias.

We hear skeptical questions from justices but explain them away as devil's advocacy or intellectual throat-clearing. The AI carries none of that emotional or ideological weight. Its dispassionate logic simply processed the available data: the justices' actual words, their voting patterns, their documented judicial philosophies. In a legal and political culture increasingly driven by partisan wishful thinking, this tool offers is an uncomfortable dose of objectivity.