Study finds signs of AI authorship in federal appellate opinions

A guest post on the Volokh Conspiracy, the legal blog hosted by Reason, presented original research suggesting a small but measurable number of US federal appeals-court opinions carry detectable traces of AI-generated text. Its author, Josh Morrow, a partner at the law firm Lehotsky Cohn, ran roughly 2,250 published circuit-court opinions issued between January and early August 2026 through Pangram, a commercial AI-detection tool, and reported that more than 50 were flagged with signs of AI authorship.

Pangram’s estimated share of AI-generated text in the flagged opinions ranged from under 1% to more than 50%, clustering toward the low end. As a baseline, Morrow ran the same tool against more than 300 opinions from January 2022, before ChatGPT’s public release, and found none flagged.

Morrow was explicit about the method’s limits. Detection tools can only read the words that make it into a final, published opinion, so they cannot see AI used for legal research, outlining, or an early draft that a judge or clerk substantially rewrote, and they cannot establish how much human review a flagged passage received. He noted that automated detection is “neither foolproof nor comprehensive,” and that his sample covered only published opinions from regional circuits — a small fraction of federal judicial output, which is dominated by district-court rulings and unpublished decisions the analysis did not touch.

The finding landed the same week Pangram was separately used to identify undisclosed AI authorship in a Wall Street Journal opinion piece by the investor Stanley Druckenmiller, one of two disclosures in quick succession pointing to the same open question: how much of the professional writing produced by people with institutional authority — judges as much as pundits — was, by 2026, being drafted with help the reader had no way of detecting.