Court tallies of AI errors show which firms get caught, not how much AI each firm size actually uses.

This explores whether anyone has measured how many AI-assisted court filings come from firms of different sizes. The corpus can't answer that directly, and the reason is worth knowing. The closest evidence is Pfefferkorn's tally of 114 US court cases with suspected AI errors. Solo practitioners or small firms were involved in 90 percent of them, and plaintiff's counsel in 56 percent Do small law firms misuse AI more often than large ones?. Those are detected incidents, though: filings where a fabricated citation or similar error was bad enough for a judge or opposing counsel to notice. They tell you who gets caught. They don't tell you who uses AI, or how much.

The gap matters because the two numbers can come apart. A large firm might use AI heavily and still rarely appear in error tallies, because several people review a filing before it goes out. Research on how organizations adopt AI points the same way. Firms that are more exposed to AI replace outside labor with AI tools faster and more cheaply than other firms, which suggests that in-house capability builds on itself rather than spreading evenly Do firms substitute labor for AI at different rates?. If law firms follow that pattern, big firms could be producing more AI-assisted work while appearing less often in the sanctions record. That idea goes beyond what the corpus measures, but nothing in the corpus rules it out.

The same problem shows up in every field where people try to count AI use. Pangram's scan of ICLR estimates that 21 percent of reviews were fully AI-generated and over half had some AI involvement How much AI content appears in peer review at ICLR?. A study of graduate admissions found that most 2025 applicants submitted essays that were likely mostly AI-written, despite a ban Does AI essay use hurt admissions chances despite quality gains?. Both numbers come from detectors run over a whole corpus, and that is exactly the method nobody has applied to court dockets sorted by firm size. Detectors also have blind spots. Graphite's report that AI-written articles leveled off at 50 percent of new articles can't separate a real plateau from newer models slipping past detection Why did AI article share stop growing after 2025?.

There is one further limit on counting. Writing-process data, such as the timing of edits and when text appears, can reliably flag wholesale delegation, where large blocks of text arrive at once. Lighter collaborative AI help looks the same as mostly unassisted work Can process data distinguish AI delegation from ordinary collaboration?. So even a well-designed study of legal filings would mostly capture the heavy-delegation end. That end is the one most likely to produce the hallucinated citations that put small firms in Pfefferkorn's sample.

In short: the corpus says small, plaintiff-side practices dominate the AI mistakes courts notice. It doesn't say how many AI-assisted filings each firm size produces. Answering that would require running an AI detector across court dockets, and no such study is in the collection.