Do small law firms misuse AI more often than large ones?

Riana Pfefferkorn counts 114 US court cases from Damien Charlotin's AI Hallucination Cases database, each one a filing that opposing counsel or the court found to contain inaccuracies "suspected or shown to have been caused by the use of AI." In that sample, 90 percent of the firms involved are solo practices or small firms (solos 50.4 percent, firms of 2 to 25 lawyers 39.5 percent), and the hallucinations are attributed to plaintiff's counsel in 56.1 percent of cases against 30.7 percent for defendants. Of the 34 cases that name a tool, ChatGPT appears in 18. Pfefferkorn presents this as confirming what "many lawyers and judges may have suspected": the archetype is "a small or solo law practice using ChatGPT in a plaintiff's-side representation."

Her explanation is about capacity more than the tool. Lawyers "struggle with busy caseloads, the trustworthiness of their co-counsel, junior attorneys, and support staff," and personal demands compete with work, and "that was already true long before AI." AI "merely provided a new way to make those errors" while "promising a way out of the underlying issues" such as time crunches and thin support. The second half of her argument comes from a Stanford study she cites, which found "disturbingly high levels of hallucinations in all the tools they studied." She places the costs on clients, opposing counsel and courts, not only on the lawyers who are sanctioned.

The vendor side of this is the study in How often do legal AI tools actually hallucinate citations?. Pfefferkorn notes that the companies behind the four products it tested (GPT-4, Lexis+ AI, Westlaw's AI-Assisted Research and Ask Practical Law AI) all appear among the 34 named-tool cases. That connects a measured failure rate to a courtroom record, though the excerpt does not say which product was used in which filing. The admissions study is a useful contrast: Does AI essay use hurt admissions chances despite quality gains? compares flagged users with comparable non-users, whereas this count has no comparison group at all. The problem of what a detection channel can see also recurs in Why did AI article share stop growing after 2025?: a database holds only the incidents someone found.

What the excerpt does not establish is a denominator. The 114 cases are incidents found in one database, covering US courts from June 2023 to October 7, 2025, with pro se matters excluded. There is no count of how many filings each firm type produces, so "90 percent" is a share of caught cases, not a rate of AI misuse by firm size. If large firms have their errors caught or reported less often, the sample understates them, and the excerpt cannot tell. Solo status and firm size were partly guessed from third-party reporting that Pfefferkorn flags as possibly outdated, and the tool breakdown covers only 34 of 114 cases (30 percent). The defensible reading is that detected AI-tainted filings cluster in small, plaintiff-side practice. Claims that small firms use AI more, or err more per filing, would need a base of AI-assisted filings by firm size that this sample does not provide.