Indicators of AI Authorship in Federal Appellate Opinions

I am pleased to share a highly engaging analysis conducted by a partner at a prominent law firm, which examines the intersection of AI and judicial writing.

The author was intrigued by what looked like signs of AI authorship in a recently issued appellate ruling from a federal court. To test this hunch, the opinion was submitted to Pangram, a modern AI-detection platform designed to separate human prose from machine-generated text. The tool’s output corroborated the hunch, marking several passages—some extending across pages—as likely AI-produced.

Following that, the entire set of this year’s published opinions from the regional courts of appeal—roughly 2,250 documents in total—was processed through Pangram, and dozens exhibited indicators of AI-generated writing.

There is no intention to identify the individual judges responsible for the opinions. When used judiciously, AI can sharpen both reasoning and expression, and it is heartening that some courts appear to be experimenting with it. For now, the aim is to outline the observations, interpret their meaning, and offer reflections on how courts can harness AI without surrendering human judgment.

Foundational Context

AI-authored text can reveal “tells,” but no single checklist captures all such signals. Sometimes a passage simply gives the impression of AI authorship. This sense arose while examining a regional-court opinion. (The observation is not tied to any particular practice area.)

I uploaded the ruling to Pangram, a cutting-edge detector trained to distinguish human writing from AI-generated prose. There remains a perception that AI writing is impervious to detection. Earlier tools lacking reliability fed that notion, but that is no longer the case. Some AI developers are building their own detectors, while third-party tools have made substantial progress. Pangram, for instance, advertises a “false positive” rate of merely 0.0041%, or about one false alarm for roughly 24,000 documents. In my experience, that claim aligns with what I observed.

Across hundreds of documents with clear authorship, I have never witnessed Pangram produce a false positive. This shouldn’t surprise us. Readers frequently recognize AI writing when they encounter it, so a detector trained to separate human from AI prose is bound to identify recurring stylistic patterns.

The results for this single opinion were noteworthy: Pangram categorized large swaths of text as AI-written. It even displayed its reasoning, labeling discrete segments in the opinion as human or AI, and attributing numerous passages—some spanning hundreds of words—to a non-human author.

What the findings suggest

To determine whether this one opinion was an anomaly, I expanded the analysis. With Pangram’s assistance, I gathered and tested the full text of every published opinion I could access from the regional courts of appeals between January and early August 2026—approximately 2,250 opinions. I also sought a baseline: what Pangram would say about opinions from the pre-AI era?

As a baseline, I ran Pangram on all published circuit opinions from January 2022—over 300 documents. Pangram reported no indicators of AI-generated text in any of them. Not a single passage, not a single sentence. Every opinion returned “0.000000% AI.” That result gave me confidence that Pangram’s classifications were not mistaking conventional judicial prose for AI authorship. (Pangram’s methods are date-insensitive, so timing did not influence these results.)

By 2026, the results looked significantly different. More than 50 opinions showed signs of AI authorship, with percentages ranging from under 1% to more than 50% AI-written, though most results clustered toward the lower end. The contrast with January 2022 was striking.

Of course, this was not an exhaustive study. I focused solely on the published opinions from the regional courts of appeals this year. Including unpublished circuit decisions would have expanded the corpus, and adding district courts would have broadened it further still. This effort did not aim to produce judge- or court-specific conclusions.

Rather, the goal was to gauge whether appellate courts are employing AI at all to assist in drafting opinions. Pangram generously provided access to enable this inquiry.

What this means

The results constitute strong evidence of AI authorship, but they are not definitive. Pangram is robust, but automated detection remains neither foolproof nor all-encompassing.

For opinions showing substantial AI-writing indications, the author rechecked them in a separate session. The findings remained consistent, dispelling concerns about random fluctuations or instability affecting Pangram’s classifications. Pangram also reported zero AI-generated prose in the January 2022 opinions. It would be highly improbable for the AI-era uptick to arise from a cause unrelated to AI.

There are also reasons to think the evidence understates how often chambers use AI. Pangram evaluated only the final opinions, whereas AI can contribute to brainstorming, research, outlining, or critiquing drafts. All of these are valuable use cases, yet they leave scant trace in the finished product. And since human edits can dull the detector’s signals, drafts that receive modest human refinement would likewise show less AI influence.

Thus, what Pangram detected should be read as partial evidence of a broader shift in how some chambers draft opinions. Even if the evidence were conclusive—or if the results understated actual use—important questions would remain. Pangram’s conclusions about AI-generated prose in an opinion do not reveal how the drafting process integrated AI—whether through collaboration with clerks, editors, or others. Nor can we reconstruct any subsequent human review.

Nevertheless, the results convinced me that the core phenomenon—federal appellate courts using AI to craft opinions—is farther from hypothetical than I might have anticipated. In fact, at least two district judges have acknowledged AI-influenced drafting in their opinions. And litigants have recently claimed a similar development in state courts. Perhaps it was only a matter of time before federal appellate courts followed suit.

Why this matters

Even with limited evidence, I do not believe the phenomenon warrants ignorance or silence.

Our initial response should be to recognize the potential benefits AI can bring to thinking and writing. The technology “has great potential,” as some senior jurists have suggested. And as another jurist has argued, these tools hold substantial promise for accelerating research, clarifying complexity, and assisting judges in fulfilling the duties their roles demand.

Nor is AI drafting inherently suspect. On the contrary, today’s AI tools can significantly enhance reasoning and refine expression. So the signs of AI authorship do not, in themselves, imply that the opinions are deficient.

At the same time, an overreliance on AI drafting should be avoided. An opinion’s drafting involves more than merely presenting a conclusion. The act of writing compels the author to organize premises, identify gaps, weigh competing considerations, and explain why one argument wins over another.

In short, writing is an essential part of judging—the exercise of judgment—not merely a record of it. AI-assisted drafting need not derail that process. Yet if large portions of text emerge essentially complete and remain largely unedited, it is fair to question whether the rigor inherent in the writing has been preserved.

Evidence of AI writing in judicial opinions could also prompt questions about whether courts rely on AI for more than drafting. Using AI to shape human judgment into polished prose is one thing; employing AI to review the record or explore counterarguments can be useful too. However, delegating ultimate judgment—the what and why of a decision—would be a very different matter altogether.

I doubt any courts have reached that point. But a deed’s text alone cannot reveal the full drafting process, independent of how much of the writing may be AI-assisted. This, too, deserves discussion.

I am encouraged that dialogue on these issues has already begun. A legal scholar has argued persuasively about AI authorship in judicial opinions and its implications for the future of legal reasoning and human participation in the system. Normative questions naturally follow: when does AI drafting align with a judge’s duty? If ever, should disclosure occur? And to what extent should AI indicators affect an opinion’s legal weight?

Empirical questions linger as well: how frequently do judges or their clerks employ AI to write? how much of that writing undergoes human revision? And as AI tools improve, will AI-generated judicial prose become more common? No definitive answers are offered here. The aim is simply to show that the dozens of opinions identified elevate these questions to renewed importance.

Conclusion

Undoubtedly, AI has a role in the legal profession—likely a substantial one—and practitioners who deny this risk losing ground in the ongoing race for efficiency and expertise that shapes outcomes for clients.

At present, I am less convinced that broad swaths of unedited AI prose should occupy the precedential opinions through which the nation’s federal appellate courts explain the law governing millions of people. “Good writing … is clear thinking made visible,” as the saying goes.

None of the dozens of opinions flagged by Pangram leads me to conclude that any judges or clerks have outsourced all of their writing to AI. Yet if such a shift ever begins, courts must guard against allowing their own judgment to wither alongside it.