Marc D. Alexander
Attorney and Mediator
Alternative Resolution Centers (ARC)
Artificial Intelligence programs such as Claude, ChatGPT, and Gemini excel at summarizing an opinion, laying bare its logic and assumptions, and picking apart unsupported conclusions. Nowhere are those capabilities better displayed than when AI's capabilities are harnessed to dissect a complicated opinion.
The recent Supreme Court slip opinion about tariffs and the International Emergency Economic Powers Act, Learning Resources, Inc. v. Trump (pub. 2/20/26), provided an excellent opportunity as a test case for the use of AI. The 6-3 majority opinion held the IEEPA did not authorize the president to impose tariffs. The majority opinion, with concurrences and dissents, is 170 pages long. The syllabus alone, which does not discuss the 80 pages of dissents, is six pages. Using Claude Sonnet 4.6, we uploaded the majority opinion with concurrences and dissents and asked Claude to: Summarize the majority opinion and the dissents. Convert this court majority opinion into formal syllogisms. Do the same for the dissents. Show the logical structure of this majority opinion and do the same for the dissents. Identify logical fallacies in the majority opinion and in the dissents. Identify assumptions of the majority opinion and of the dissents. Identify unsupported assumptions of the majority opinion and the dissents. Despite our poorly worded reference to "dissents," Claude helpfully included responses for concurrences.
The response to our questions, obtained in two minutes once the majority opinion with concurrences and dissents had been uploaded, was thorough and 18 pages long. The AI logic is implacable and unemotional. Unless we provide additional instructions, Claude does not care about what we think about the Justices' disagreements. Claude does not care whether we support the majority opinion or the dissents, or whether we are Republicans or Democrats. Claude replies much more quickly than is humanly possible. We could ask Claude to draft a Claude opinion after digesting the briefs, but we did not. That opinion would be reasoned and logical, and we could interrogate Claude to list assumptions upon which it constructed the opinion.
For the tariff case, Claude generated 11 syllogisms, covering the logic of the majority opinion, the concurrences, and the dissents. Each syllogism was in the form: "P1 = First Premise (a proposition asserted as true); P2 = Second Premise (a second proposition asserted as true); C = Conclusion (what logically follows if both premises are true)."
We'll provide three examples of Claude's syllogistic reasoning. The first syllogism, covering the majority opinion drafted by Justice John G. Roberts and joined by Justices Sonia Sotomayor, Elena Kagan, Neil M. Gorsuch, Amy Coney Barrett, and Ketanji Brown Jackson, and relying on a textual reading, is: "P1: Congress may only delegate a power it has used clear, express language to convey; the power to impose tariffs (a branch of the taxing power) is constitutionally Congress's alone and requires express delegation. P2: IEEPA delegates authority to "regulate...importation" without mentioning tariffs, duties, or any revenue-raising power; no U.S. statute uses "regulate" to grant taxing authority. C: IEEPA does not delegate tariff authority to the President."
Roberts, Gorsuch and Barrett, while joining in the majority opinion, additionally relied on the Major Questions Doctrine to join the majority. This required different logic, and therefore Claude generated Syllogism Two: "Under the major questions doctrine, when the executive claims extraordinary delegated power of enormous economic and political significance, it must point to clear congressional authorization. Tariff power over the entire American economy without procedural limits is the paradigm case for the doctrine. No prior president ever invoked IEEPA for tariffs. The fact that the stakes "dwarf" all prior MQD cases reinforces caution. No exception applies for emergency statutes or foreign affairs."
Justice Kagan did not rely on the Major Questions Doctrine and made this clear in a concurring opinion. This is summarized by Claude with Syllogism 5: "P1: Proper statutory interpretation requires reading the specific delegation text in light of its surrounding statutory context, Congress's pattern of usage, and common sense about how Congress delegates. P2: "Regulate" does not mean "tax" in ordinary usage or in any other statute; Congress always uses distinct vocabulary for tariff power; no President before 2025 read IEEPA to authorize tariffs; reading tariff power into IEEPA would eviscerate all the constrained tariff statutes in Title 19. C: IEEPA does not authorize tariffs under ordinary statutory interpretation, without need of any special doctrine."
Claude also provided the syllogistic reasoning for the Jackson concurrence and the Clarence Thomas, Brett Kavanaugh, Samuel A. Alito dissents.
Claude described the majority's reasoning as "a layered opinion," meaning that six of the justices arrived at the same result through independent legal rationales. The first layer relied on the constitutional proposition that the tariff power belongs to Congress, and that the president had no inherent power to impose a tariff. (The defense conceded in oral argument that the president did not have inherent power to impose a tariff.) Therefore, he had to rely entirely on whatever power had been delegated through the IEEPA. The second layer, covering Justices Roberts, Barrett and Gorsuch, relied on the proposition that the tariff triggered the Major Questions Doctrine, and Congress had not clearly delegated the tariff power under the IEEPA. The third layer relied on textual analysis that ordinarily "regulate" does not mean "tax." The benefit of this summary is that it instantly shows the second layer relying on the Major Questions Doctrine reasoning is dictum, because the majority opinion requires additional constitutional and textual analysis.
In addition to spelling out the logical syllogisms of the majority, concurrence, and dissent opinions, Claude can also identify logical fallacies, stated assumptions, unstated assumptions and unsupported assumptions. We give a few of the examples that Claude generated.
An example of a majority opinion logical fallacy is generalization from congressional practice. The majority argued from the explicit use of terms such as duty and surcharge when Congress delegated its authority in ordinary tariff statutes, that the absence of such terms in an emergency statutory context was compelling evidence of a lack of congressional delegation. While this may be rhetorically persuasive, it is not tight logic, for Congress could reasonably use different language in an emergency situation. One could find the same point made in Justice Kavanaugh's dissent, though two minutes with Claude gets to this point.
An example of a logical fallacy can be found in Thomas's dissent, which includes an anachronistic historical argument. Thomas argued that in colonial times, tariffs belonged to the royal prerogative of the British King. However, relying on pre-Revolutionary royal authority to support the president's tariff authority is questionable, especially since the Revolutionary War questioned the king's right to impose tariffs on the American people. Again, one could come to the same conclusion in the fullness of time by reading Gorsuch's historical analysis of the royal prerogative and the role of Parliament.
A Claude example of an explicit majority assumption is that taxing and regulatory powers are categorically distinct. An unstated assumption is that a 1979 definition of "regulate" in Black's Law Dictionary accurately captures what Congress meant in 1977 in the IEEPA. An example of an unsupported majority assumption, perhaps better described as an inferential leap, is that non-use of the IEEPA as a tariff statute for 47 years meant that Congress did not intend to authorize the IEEPA to be used for tariff purposes.
Claude's example of an explicit dissent assumption in the Kavanaugh dissent is that "regulate ... importation" in ordinary meaning encompasses tariffs. An unstated assumption in the Kavanaugh dissent is that the absence of the word "tariff" or "duty" in the IEEPA is explained by the statute's focus on emergency situations, and not as intent to exclude tariffs from the scope of the IEEPA. An example of an unsupported dissent assumption is that Federal Energy Administration v. Algonquin, 426 U.S. 548 (1976) which Kavanaugh used to justify the exercise of the power of delegation of tariffs to the president, is "on all fours." Now actually "on all fours" was not Kavanaugh's language, but rather Claude's shorthand. However, Kavanaugh did say that Learning Resources "should follow a fortiori from Algonquin," and that "[n]o meaningful daylight exists" between the statutory phrase in Algonquin and that in Learning Resources. Perhaps Claude should have quoted exact language, but the phrase "on all fours" did not overreach. However, Algonquin also referred to broad "take such action as he deems necessary" language and to "duty," unlike the language of the IEEPA. While the facts are similar, Algonquin is not controlling.
We draw a few conclusions from this experiment in which we turned the power of a current AI program on a complex opinion with concurrences and dissents. First, the results are provided with remarkable speed and detail. Second, we could achieve similar critiques of the opinions by reading and analyzing 170 pages of text, but we are not confident that the results would have been more accurate or complete. We did review the majority opinion, concurrences and dissents to see if Claude's points were accurate. However, if we wanted to rely on those results in a legal brief, we would need to spend more time reviewing the opinion, concurrences and dissents. Third, one reason the results seem to be coherent is that Claude reviewed a single document with multiple opinions, and we did not ask it to search the web for additional information. Thus, Claude worked with limited and self-contained information. Claude could make use of the majority opinions, concurrences and dissents to criticize one another; in other words, the opinions had already done some of the destructive criticism we asked Claude to do for us. However, Claude's additional contribution is its usefulness showing the logical structure of the opinions and their assumptions.
One more conclusion is disquieting. It is sometimes said that no matter what one's preconceptions may be, one can always find something on the internet to support one's own opinions and one's desired result. The dizzying speed and ease with which Claude can spin out logic, fallacies, explicit assumptions, unstated assumptions and unsupported assumptions may mean that with the use of AI, advocates and judges will usually be able to find arguments to support their conclusions. Those arguments may be unconvincing, but they will lead to a logical conclusion.
On the other hand, the speed and ease with which Claude can reveal the assumptions that underlie an argument may help advocates and judges cut to the chase. Complex legal arguments are often grounded on a few assumptions that AI can quickly lay bare. That was the case here.