How AI Could Have Improved Justice Sotomayor's Recent Dissent

A case study in collaborative legal critique

On May 11, 2026, the Supreme Court issued a brief order in Allen v. Caster vacating a District Court judgment that had enjoined Alabama's 2023 congressional redistricting plan. Justice Sotomayor, joined by Justices Kagan and Jackson, filed a dissent. The dissent is forceful, but I think it makes a strategic choice that weakens what could have been a much stronger opinion. This post is about how I worked through that problem with Claude, and what the exercise suggests about the future of judicial drafting. The short version is that the liberals on the bench could have issued a more forceful opinion if they'd proceeded with an open mind and invited AI into the conversation. You can see the "AI-improved" dissent in the gray box at the bottom of this post.

The background, in brief

For readers who don't follow voting rights litigation closely, a few words of context are necessary. It may still be incomprehensible to the uninitiated but my hope is that it gives such readers a fighting chance.

The Voting Rights Act of 1965 contains two operative provisions for present purposes. Section 5, which required certain jurisdictions to get federal preclearance before changing voting rules, was effectively nullified by Shelby County v. Holder in 2013. Section 2, which (since 1982) prohibits voting practices that "result in" racial discrimination, remained the principal tool for challenging discriminatory districting. Under the framework as interpreted in Thornburg v. Gingles (1986), a plaintiff alleging that a districting plan "dilutes" minority votes had to show three things: that the minority group was large and compact enough to form a majority in a reasonably drawn district, that the minority voted cohesively, and that the white majority voted as a bloc to defeat minority-preferred candidates. If those preconditions were met, a court would assess the "totality of the circumstances" to decide whether the plan denied minority voters equal opportunity to elect representatives of their choice. At least for a while, section 5 as interpreted in Gingles resulted in Congressional maps in which minority voters – and derivatively Democrats – had a much greater opportunity to elect their preferred candidates.

In 2023, the Supreme Court applied that framework in Allen v. Milligan and held that Alabama's congressional map violated Section 2 by including only one majority-Black district where the geography and demographics could have supported two. The Court ordered Alabama to draw a remedial map with a second opportunity district.

Alabama refused. It drew a new map—the 2023 Plan—that still contained only one majority-Black district. After an 11-day trial, a three-judge District Court issued a 268-page opinion in a case captioned Allen v. Caster, finding that Alabama had not merely failed to remedy the prior violation, but had affirmatively engineered new districting criteria designed to make a second opportunity district mathematically impossible. The court found procedural irregularities, manufactured legislative findings, and pretextual criteria invented for the occasion. It concluded that Alabama had violated the Fourteenth Amendment through intentional racial discrimination, and it enjoined the 2023 Plan.

Then, on April 29, 2026, the Supreme Court decided Louisiana v. Callais, and substantially rewrote Gingles. The Court reasoned that the old framework was on a collision course with the colorblindness the Fourteenth Amendment requires, and that legislation enacted under Section 2 of the Fifteenth Amendment had to be "congruent and proportional" to the constitutional violation it remedied. From those premises, the Court derived a series of new requirements. Section 2 plaintiffs must now show that their illustrative maps satisfy all of the State's stated political goals (including partisan gerrymandering objectives), must control for partisan preference when proving racially polarized voting, and must focus the totality-of-circumstances inquiry on present-day intentional discrimination rather than historical patterns. The practical effect, as Justice Kagan explained in her Callais dissent, is to convert Section 2 from an effects test into something close to an intent test. The majority in effect had to destroy the statute in order to save it. It is a blockbuster decision whose sequelae may well determine whether the Democrats can regain a majority in the House of Representatives following the 2026 midterms.

And today, just twelve days later, the Court vacated the Alabama judgment in Allen v. Caster and remanded "for further consideration in light of Callais." That is the order Justice Sotomayor dissented from.

The dissent's problem

Justice Sotomayor's dissent makes two main arguments. The first, and the one she leads with, is maximalist: Callais is entirely irrelevant to the Alabama case because Callais construed Section 2 while the District Court's finding rested on the Fourteenth Amendment. The second, tucked in near the end and introduced with the words "if the Court must act now, it should at least," is that the intentional-discrimination finding should be affirmed under clear-error review because the District Court's factual findings are overwhelming. A stronger version of the dissent would have inverted the order and built the opinion around a concrete affirmative ask: a limited remand on Section 2 in light of Callais, with the Fourteenth Amendment injunction left intact pending the District Court's reconsideration. That disposition would have rested on the same clear-error point but reframed it as supporting a specific remedy rather than as a fallback objection.

The second argument is much stronger than the first. The first argument invites the obvious response that Alabama's conduct was framed as defiance of a federal decree, and if Callais has narrowed what that decree could properly have required, then the inference of racial motive is at least partially weakened. A justice writing in response could draft something like this:

A State that declines to do what federal law does not require has not, by that refusal alone, betrayed a racial motive.

That sentence does real damage to the maximalist position, because it operates at a level of abstraction where it sounds unanswerable. The only way to defeat it is to refuse the abstraction and force the argument back onto the record—to show that Alabama did not merely refuse, but acted affirmatively, through pretext and procedural irregularity, in ways that violate the Fourteenth Amendment regardless of what Section 2 demands. That move depends on the District Court's factual findings, and those findings are protected by clear-error review. Which is exactly why the dissent's buried argument is the stronger one: it cannot be answered by abstraction because it does not operate at the level of abstraction.

Working through it with Claude

I worked through this with Claude across several exchanges. The process was iterative and, I think, instructive.

I started by asking Claude to draft a hypothetical Alito concurrence responding to Sotomayor's dissent. This was useful as a stress test. Claude produced something that I described at first glance as being "frankly, pretty telling." The hypothetical concurrence pressed exactly the abstraction problem above: it framed Alabama's conduct as a "refusal to do what federal law did not require," which made the intent finding look derivative of the Section 2 ruling.

I then asked Claude to draft Sotomayor's response to that imaginary concurrence. Claude's response identified the right move—the "refusal alone" framing is not a description of this case, because the District Court found affirmative manipulation, not mere inaction. But the exercise revealed something important: Sotomayor's actual dissent had not made that move with the clarity it deserved. The response I asked Claude to imagine was, in effect, the dissent Sotomayor should have written.

The next step was the most useful. I asked Claude who had the better argument, with the constraint that we had to accept Callais as binding even if we thought it was wrongly decided. This constraint matters. A dissent that rejects the controlling precedent will be dismissed as resistance. A dissent that works within the precedent and shows its limits is much harder to answer. Claude's analysis—that Sotomayor wins on balance but that her own framing makes the win harder than it should be—matched my own reaction, but with sharper articulation of why.

I then asked the question that produced this post: would the dissent have been stronger if Sotomayor had pressed only her alternative argument? Claude's answer was unequivocal yes, and the reasoning was about opinion craft rather than doctrine. Maximalist framings invite the strongest possible counterresponse. Disciplined framings deny the counterresponse purchase on the actual case. The phrase "at least," used to introduce the clear-error argument, is what Claude called "the tell"—it signals concession rather than confidence, and tells the reader this is the fallback.

Finally, I asked Claude to draft an improved dissent that could not respond to the hypothetical Alito concurrence (because that concurrence is imaginary) but had to anticipate the strongest counterarguments and address them within the structure of the opinion itself. The result opens by naming a concrete alternative disposition the Court could have chosen—a limited remand on Section 2 with the Fourteenth Amendment injunction left intact pending reconsideration—and builds the entire opinion around that proposal. It leads with the standard of review and the District Court's factual findings, concedes Callais's authority over Section 2 in order to locate the constitutional finding outside its reach, addresses the "defiance" objection head-on rather than ignoring it, and frames the Court's actual disposition as aggressive rather than cautious by showing how easily a narrower disposition would have accommodated every legitimate concern. I've added Claude's opinion as an appendix to this post.

What the collaboration actually looked like

I want to be clear about the division of labor, because I think it matters for understanding what AI can and perhaps cannot contribute to this kind of work.

I did not permit Claude initially to identify the strategic flaw in the dissent on its own. I came to that view through my own reading of the imaginary opinions, and I directed the exchange toward testing it. What Claude contributed was rigorous adversarial pressure on each step of the argument. When I proposed that Sotomayor's maximalist framing was overreach, Claude did not simply agree. It first wrote the strongest possible Alito response, then wrote Sotomayor's strongest possible counter, then assessed both with the constraint of accepting Callais. The discipline of forcing Claude to argue each side in turn produced sharper analysis than I would have generated working alone.

The drafting itself was genuinely collaborative. I directed structure and emphasis. Claude executed prose and identified counterarguments I had not surfaced. When the improved dissent addresses the "defiance" objection directly—conceding that the argument has "surface appeal" before showing why it fails on the record—that move was Claude's. I would have ignored the objection, which would have made the dissent weaker. Claude's instinct to address the strongest counterargument rather than avoid it is, I think, a genuinely useful contribution to legal drafting.

A suggestion

The dissent reflects a strategic choice, and one can imagine reasons for it: a desire to make a broader doctrinal point about the relationship between Section 2 and the Fourteenth Amendment, a calculation that the maximalist framing better serves future litigation, perhaps even deep anger at a Callais decision that threatens to cement the incumbency of a political party that generally despises her ideology and that of her liberal colleagues. But on the merits of this particular opinion, in this particular procedural posture, the choice was unimpressive. The clearest evidence is what the majority did not do: six justices vacated the District Court's judgment without writing a word in response. No concurrence, no engagement, no felt need to answer. That silence is not proof of the dissent's weakness—the press of a late term and the shadow-docket posture both counsel brevity. Moreover, the silence could just be a contemptuous flex by those whose power now goes unchecked. But it is suggestive. A dissent that had pressed the clear-error point as its lead argument, and had forced the majority to confront a 268-page factual record reviewed under the most deferential standard in our jurisprudence, would have been harder to ignore. The majority might still have vacated. The Callais-inspired rush to partisan gerrymander from sea to shining sea would likely continue unabated. But perhaps a more telling dissent would have impelled the majority to at least engage.

What the collaboration with Claude demonstrates is that just as AI can help even seasoned Supreme Court advocates like Neal Katyal, an adversarial AI interlocutor can stress-test a justice's draft in ways that even a team of a justice's clerks can not. An AI likely lacks loyalty to a particular justice's ideology. Asked to do so, it will write the strongest possible counterargument, then write the response, then assess both honestly when constrained to do so. That is a different kind of useful than what a clerk provides, particularly where justices select them in part for ideological compatibility, and I see no reason justices should not have access to it.

I predict that the Supreme Court will not be the first institution to adopt AI in this manner. I further predict that it will not admit doing so for a very long time. District courts and circuit courts will, however, I suspect. Solicitors general and appellate advocates already are. But the highest court in the country, whose opinions shape the law for decades, would actually benefit from the same adversarial testing that the rest of the legal profession is beginning to take for granted. The dissent in Allen v. Caster should have been the dissent Claude and I produced in an afternoon. That it was not is a small loss for the Court and a larger loss for the litigants and voters whose interests the dissent was meant to defend.

I am not (quite) suggesting AI should (yet) write Supreme Court opinions. I am not suggesting that a Houston law professor knows better than a Supreme Court justice what to say. I am suggesting, however, that the gain in persuasive power created by combining knowledgeable voices with a powerful AI is telling. Modern AI has gotten good enough to deserve a seat in the room when important opinions are drafted.

Afterthought 1: A Preemptive Response to Criticisms

After I read this post, I recognized that it might be controversial. So I decided to preemptively tackle at least a few of the anticipated blasts.

First, on coalition management. A Supreme Court Justice writing in dissent is not just trying to win the rhetorical contest of the moment. She is writing for a bloc that has to hold together, for future litigation that has to be preserved, for the historical record, and for the press. What looks from outside like a craft failure may be a deliberate choice to keep a broader doctrinal argument alive even at the cost of immediate persuasive power. The maximalist framing I criticize—that Callais is wholly irrelevant to the Fourteenth Amendment finding—may serve future cases in ways the narrower clear-error argument would not. This post is about craft narrowly construed. It is not a claim that I understand the strategic calculus better than the chambers that made it.

Second, on hallucination. The objection that AI cannot be trusted in legal writing because it invents citations is serious, and it would be disqualifying if true of this exercise. It is not. Every citation in the improved dissent was verified through Midpage, the legal research tool I rely on for citation accuracy. The verification took less time than the drafting did. The lesson is not that AI is reliable on its own. It is that AI plus a competent lawyer doing verification work produces output that is at least as accurate as a draft produced by clerks alone, and considerably faster. Responsibility for accuracy remains with the lawyer. That is exactly where it should be.

Third, on loyalty. I do not claim AI is unbiased. Claude has been trained by a company with its own commercial and ideological incentives, and those incentives shape its outputs. The point is narrower. Modern AI should be able to be directed, on demand, to argue any side of a question with comparable rigor. That is a different kind of useful than what an excellent clerk provides. Clerks are excellent. AI is differently excellent. The two are complements, not substitutes.

Afterthought 2: What this means for legal education

The narrow lesson here is about one dissent. The broader lesson is about teaching. Most discussion of AI in legal education focuses on summarizing cases, finding authorities, drafting memos, and editing prose. What I did with Claude on the Sotomayor dissent was something else: I used the model adversarially. I asked it to generate the strongest attack on an argument, then the strongest response, then to assess both under the constraint that Callais binds. That sequence resembles what a good moot-court judge or skeptical appellate partner does. It does not resemble what a research assistant does.

Legal education has long treated strategic judgment as something students absorb after enough Socratic questioning. The exchanges I had with Claude suggest a more explicit approach. Which argument should lead? Which should be conceded? Which invites a devastating response? Which constraint changes everything? A student can practice these questions directly, with an interlocutor that will argue any side on demand.

The pedagogy follows. Students should learn to argue under constraint—to accept a hostile precedent as binding and ask what remains, rather than denouncing the precedent and changing the subject. Faculty should use AI to generate alternative structures of an argument and then ask students to compare them. Assignments should shift from "write a dissent" to "submit your initial theory, the strongest AI-generated attack on it, your revised theory, and an explanation of what you changed and why."

The danger is not just hallucinated citations. It is fluent strategic nonsense—an AI argument that sounds elegant but collapses on contact with the record. That risk does not defeat the pedagogical use. It defines it. AI output is not authority. The question is never whether the AI wrote well but whether what it wrote is true, supported, strategically sound, and appropriately constrained.

The shift I am describing is from product to process. Old assignments asked for the memo, the brief, the rule statement, the exam answer. The new ones ask how the student got there—what alternatives were considered, what counterarguments were tested, what was verified, what was rejected, and why this version is better than the plausible others. The interesting thing about my exchanges with Claude was not that the model could draft an improved dissent. It was that the back-and-forth showed why one argumentative structure beat another. That is what teaching is supposed to do.