Signs of AI Authorship in Federal Appellate Opinions

I was intrigued to see what struck me as indicia of AI writing in a recent, published opinion from an Article III appellate court. So I took the opinion to Pangram, an AI-detection tool. Pangram’s results backed my intuition, identifying multiple passages, some stretching for pages, as AI-generated.

Then I ran the full set of this year’s published opinions from the regional courts of appeals through Pangram—about 2,250 opinions in all. Dozens show signs of AI writing.

No, I’m not going to name the judges who wrote or joined the opinions. Used well, AI can sharpen both thinking and prose, and I am heartened that some courts appear to be using it. For now, I just want to survey my findings, explain what they mean, and offer a few thoughts about how courts can make the most of AI without ceding human judgment.

Background

AI writing can leave “tells,” but the full set resists reduction to a checklist. Sometimes you just get the feeling a passage is AI. That feeling struck recently as I was reading an opinion from one of the regional courts of appeals. (Like the other opinions I discuss, this one has no connection to my practice.)

I uploaded the opinion to Pangram, which is a cutting-edge AI-detection tool trained to distinguish human writing from AI. There is a lingering perception that AI writing is invulnerable to detection. Earlier tools touting this ability were faltering and gave that view some basis. But no more. Some AI developers are working toward offering their own proprietary detectors. Meanwhile, third-party detectors have improved considerably. Pangram, for example, claims a “false positive” rate “of just 0.0041%, or roughly one false positive for every 24,000 documents.” My experience is in accord.

Across hundreds of documents where I’m certain of authorship, I have never seen Pangram give a false positive. This should not come as any shock. After all, many readers can discern AI writing when they encounter it. It seems perfectly natural that a tool trained to distinguish human writing from AI writing could see some of the same patterns.

Pangram’s results for this single opinion were notable: it classified swaths of text as AI-written. Pangram showed its work too, labeling discrete segments in the opinion as either human or AI. It attributed numerous passages—some stretching for hundreds of words—to a non-human author.

What I found

Curious whether this sole opinion was an outlier, I broadened the search. With help from Pangram, I assembled and tested the full text of every published opinion I could readily obtain from the regional courts of appeals from January through early August 2026—about 2,250 opinions in all. I also wanted a baseline: what would Pangram say about opinions from the pre-AI era?

To get a baseline, I tested all the published circuit opinions from January 2022, more than 300. Pangram found no signs of AI-generated text in any of them. Not one passage, and not one sentence. Every opinion returned “0.000000%” AI. That gave me some confidence that Pangram was not mistaking the ordinary conventions of judicial writing for AI authorship. (Pangram’s methods are insensitive to dates, so I did not worry that the opinions’ timing influenced these results.)

The results for 2026 looked very different. More than 50 opinions showed signs of AI authorship. The exact percentages ranged from less than 1% to more than 50% AI-written, with most results clustering near the bottom of that range. Still, the contrast with January 2022 was hard to miss.

Of course, this was not an exhaustive study. I looked only at this year’s published opinions from the regional courts of appeals. Adding unpublished circuit decisions would have multiplied the corpus, and adding district courts would have made it enormous. Nor was I looking for judge- or court-specific results.

Instead, I was interested in evidence bearing on whether the circuit courts are using AI at all to help draft opinions. Pangram generously supplied credits to make the search possible.

What it means

The results are strong evidence of AI authorship, but they are not conclusive. Pangram is robust, but automated detection is neither foolproof nor comprehensive.

For opinions with substantial signs of AI writing, I scanned them again from a separate account. The results were consistent with the first pass, thus dispelling my concerns about random noise or instability influencing Pangram’s classifications. Pangram also flagged zero published circuit opinions from January 2022 as containing AI-generated prose. It would be a remarkable coincidence if the AI-era uptick owed to some cause other than AI.

I had more subjective reasons for crediting the results, too. Pangram’s methodology, which I will not detail here, is sensible: broadly speaking, it uses a classifier trained on human and AI writing to distinguish one from the other. Said differently, Pangram takes some of the same techniques that make AI so powerful and uses them to flag AI prose. Then there is the prose itself. Many of the passages Pangram flagged exhibited the stylistic tells that prompted my interest in the first place, and very few of the results left me incredulous at Pangram’s attributions.

There are also reasons to think this evidence understates how often chambers are using AI. Pangram analyzed only the final opinions, but AI can also brainstorm, research, outline, or critique a draft. All are great use cases, but they leave little evidence of AI assistance in the final product. And because human editing can weaken the signals that Pangram relies on, the same goes for drafts that see modest human editing.

What Pangram detected, then, is best understood as some evidence indicating a broader shift in how some chambers are writing opinions. Even if the evidence were conclusive, or if the results did understate actual usage, important questions would persist. Pangram’s conclusions about AI-generated prose in an opinion do not tell us how the language entered the drafting process, which often unfolds as a collaboration between judges and law clerks. Nor can we reconstruct whatever human review followed.

Still, the results convinced me that the core phenomenon—federal appellate chambers using AI to draft opinions—is a few steps further from hypothetical than I might have otherwise guessed. Indeed, at least two district judges have already acknowledged that AI writing entered their opinions. And litigants recently claimed that the same occurred in state court. Perhaps it was only a matter of time before federal appellate courts followed suit.

Why it matters

Even though the evidence is limited, I am not convinced the apparent phenomenon should go unnoticed or undiscussed.

Our first response should be to celebrate the salutary effects that AI can have on both thinking and writing. The technology “obviously has great potential,” as Chief Justice Roberts has argued. And as Judge Newsom has argued, these tools hold great “promise”—for accelerating research, unpacking complexities, and greatly helping judges discharge their duties with the excellence their jobs demand.

Nor do I believe AI drafting is inherently suspect. On the contrary, today’s AI tools have incredible potential to improve thinking and refine prose. So, the indicia of AI authorship I saw gave me no reason to think that the opinions are in any way deficient.

At the same time, over-reliance on AI drafting should be avoided. Writing an opinion does far more than communicate a conclusion already reached. The act of writing forces an author to organize premises, confront gaps, reconcile competing considerations, and articulate exactly why one argument prevails over another.

Writing is thus an integral part of judging—i.e., the exercise of judgment—not merely a record of it. AI-assisted drafting need not short-circuit that process. But if substantial passages are arriving essentially finished and remaining largely unedited, it is fair for the public to discuss whether any of the rigor writing itself imposes has been lost along the way.

Signs of AI writing in judicial opinions might also lead the public to wonder whether courts are turning to AI for more than just drafting. Using AI to formalize human judgment into polished prose is one thing. And using AI to help review the record or examine counterarguments can be useful too. But delegating the ultimate judgment—that is, the what and why of a decision itself—is something else entirely.

I doubt any courts have gone that far. But an opinion’s text itself ordinarily cannot tell us, no matter how much (or little) of the writing might appear to be AI-assisted. This, too, is worth discussing.

I am heartened that some of the conversation has already begun. Richard Re, for instance, has written persuasively about what AI authorship in judicial opinions could mean for “the fate of legal reason and the future of human participation in the legal system.” Normative questions naturally follow. When is AI drafting consistent with a judge’s duty? When, if ever, should it be disclosed? And to what degree should indicia of AI authorship affect an opinion’s legal value?

Empirical questions lurk too. How often are judges or their law clerks using AI to write? How much of that writing sees human revision? And as AI tools keep improving, will AI-generated judicial prose become more common? I offer no answers here. Rather, I mean only to show that the dozens of opinions I found give the questions renewed import.

Conclusion

Undoubtedly, AI has a place in the legal profession, probably a large one, and practitioners who believe otherwise risk losing ground in the battles for efficiency and expertise that drive favorable outcomes for clients.

I am presently less certain that large swaths of unedited AI prose have a place in the precedential opinions through which our Nation’s federal appellate courts explain the law governing millions of Americans. “Good writing … is clear thinking made visible,” as the aphorism goes.

None of the dozens of opinions Pangram flagged led me to conclude that any judges (or law clerks) are outsourcing all their writing to AI. But should that ever begin to occur, courts should guard against letting their judgment wilt alongside.