In 2023, a federal judge confronted a filing that cited judicial opinions no legal database could find. The lawyers in Mata v. Avianca had used ChatGPT for legal research. The chatbot supplied case names, quotations, docket numbers, and passages that looked convincingly judicial and were not real.
The court's sanctions order describes what followed: the attorneys stood by the fake opinions after their existence was questioned, supplied purported copies containing more fabricated material, and offered shifting explanations. The court imposed a $5,000 penalty and required notification of the judges falsely identified as authors.
A language model predicts plausible text. A lawyer is supposed to check whether the plausible text exists. This division of labor proved controversial.
The lawyer without lawyers
A year later, the Federal Trade Commission announced an action against DoNotPay. The company had promoted its service as the “world's first robot lawyer.” The FTC alleged DoNotPay did not test whether its service performed like a human lawyer and had not hired or retained attorneys to test most legal features. Under the proposed settlement announced in 2024, the company agreed to pay $193,000 and stop making unsupported claims.
The two episodes point in opposite directions and arrive at the same desk. In one, licensed professionals treated generated text like verified research. In the other, software was marketed with the authority of a licensed professional without evidence it deserved the title. Humans inflated the machine's role, then discovered credentials still matter after the invoice arrives.
CRR remedy
AI can help organize, draft, search, and compare. It cannot absorb professional responsibility. Verify every citation. Preserve human review. Do not sell a prediction engine as a fiduciary, doctor, lawyer, or oracle merely because “helpful text box” performed poorly in focus groups.
The robot moves that “it sounded right” be excluded from all future legal standards.
← Return to all dispatches