Make legal judgment the test for AI-assisted work

Law360 Legal Industry reported on Nov. 4 that a public database tracking cases involving hallucinations generated by artificial intelligence had recently surpassed 2,200 entries. The report also said Massachusetts District Court Associate Justice Brian Palmucci identified a deeper risk in attorneys’ use of such tools: gradually surrendering judgment to AI and losing analytical skills. The event is therefore not a report of faster drafting or cheaper review. It is a growing count of failures alongside a warning about the human capacity needed to detect them.

That matters because AI does not eliminate the consequences of bad legal work. A system may generate text at scale, but the lawyer remains exposed when the text carries a false authority, an unsupported proposition or an error that reaches a filing. Legaltech News (Law.com) describes the division plainly: AI supplies speed, scale, consistency and computational power, while attorneys supply legal knowledge, context, skepticism, judgment and accountability. The attraction of an AI-native service is unfinished unless it preserves those human functions where the work becomes legally consequential.

The database count proves that hallucination-related failures have accumulated to a substantial level. It does not prove that every failure came from the same tool, workflow or supervision model, and it does not establish that AI has already weakened the judgment of the profession as a whole. The stronger claim would require a demonstrated link between reliance on AI and a measurable decline in lawyers’ analytical performance, not only a collection of bad outcomes. Nor does the count show that human review is sufficient. A review process can exist on paper while failing to catch errors in practice. The order of responsibility remains clear: the system produces the material, but the lawyer must answer for the legal work delivered.

The warning also gives the training question a sharper edge. Reuters Legal reported on Oct. 2 that a federal judge warned that increasing use of artificial intelligence could create risks for clients and hinder the training of junior lawyers, even as the judge declined to sanction attorneys over AI-generated errors in a court filing. That decision separates consequence from punishment: the absence of a sanction in one matter does not make the work sound, and a court’s restraint does not supply the judgment that the workflow failed to develop. Automating routine work while relying on experienced lawyers to supervise it creates a real contradiction if the next generation does not acquire the same analytical skills.

The next confirming fact is not another promise that AI improves efficiency. It is a later court order or filing showing whether lawyers caught and corrected AI-generated errors before submission, or whether another hallucination entered a legal document despite a stated review process. A useful result would also identify who performed the review and what changed after the error. Until those facts appear, the count of cases supports a narrow conclusion: AI-assisted legal work still carries ordinary professional consequences, and the judgment needed to control them remains part of the service.