A federal judge warns law firms that unchecked artificial intelligence can harm clients and weaken the training of junior lawyers, even as he declined to sanction attorneys over citation errors linked to Harvey. The warning came from U.S. District Judge Arun Subramanian in Manhattan.

 

The dispute arose in Hill v. Foundation Media, a copyright case in the Southern District of New York. Foundation’s lawyers acknowledged that AI-assisted work produced an inaccurate citation and quotations that did not support the propositions attributed to them, and that the mistakes reached the court without adequate verification.

 

The record establishes three important facts:

  • The court ordered lead counsel to explain whether AI caused the errors.
  • The firm admitted that its verification policy was not followed.
  • The judge declined sanctions but issued a profession-wide warning.

 

Federal Judge Warns Law Firms After Citation Errors

Subramanian’s September 22 order dismissed plaintiff Lawand Hill’s copyright complaint against Foundation Media and entered judgment for the defendant. Yet the judge separately required Foundation’s lead counsel to file a sworn declaration addressing whether AI had been used and whether unchecked output caused the questionable authorities.

 

The order identified quotations attributed to a case called Piazza and a citation to Zhao. It did not assume that AI was responsible; instead, it required counsel to state the facts under penalty of perjury. That distinction kept the court’s inquiry focused on evidence rather than suspicion about the technology.

 

After reviewing the response, Subramanian declined to impose sanctions. Reuters reported that he nevertheless described the episode as a wake-up call and warned that AI-generated legal work must be checked carefully before it is submitted to a court.

 

Shapiro Arato Bach Admits Unchecked AI Use

In a September 28 declaration, attorney Cynthia Arato said Shapiro Arato Bach had used generative-AI tools designed for legal work when preparing Foundation’s response to Hill’s objection. A subsequent filing identified the provider as Harvey, according to Reuters.

 

Arato said the firm had conducted onboarding sessions, required mandatory ethics training and made employees sign an acceptable-use policy. That policy required AI output to be reviewed, validated and corrected before use or disclosure. The filing therefore describes a failure to follow an existing control, not the absence of a written rule.

 

The declaration states that the cited Zhao case did not exist and that the Piazza decision did not support the propositions for which it was cited. Counsel said the underlying legal principles were correct, but that does not cure inaccurate authorities in a filed brief.

 

Arato attributed the mistake to an unnamed lawyer who prepared the initial draft and finalized the response without checking the case citations. She apologized, characterized the incident as isolated and said the firm was implementing additional safeguards.

 

Background Reading

 

Harvey AI Citation Errors Expose a Control Gap

The incident shows why purchasing a legal-specific AI system does not transfer professional responsibility to the vendor. A firm can have training, policies and enterprise software yet still fail at the last checkpoint if the lawyer signing or filing the document does not verify every authority and quotation.

 

Legal AI can accelerate research and drafting, but fluency is not evidence that a citation exists or supports a particular sentence. Verification requires opening the authority, checking the relevant passage and confirming that the court, date, procedural posture and quotation are accurate.

 

Harvey declined to comment to Reuters. The record does not establish that the product itself malfunctioned, how the prompt was written or which product feature produced the material. The confirmed failure is narrower: AI-assisted citations entered a federal filing without the review required by the firm’s own policy.

 

Subramanian Links AI Oversight to Lawyer Training

Subramanian also expanded the issue beyond citation checking. He warned that increasing reliance on AI could hinder the development of junior lawyers, whose early assignments traditionally teach them to research authorities, structure arguments and learn through detailed review by senior attorneys.

 

The judge suggested that firms could require initial drafts of briefs to be prepared without AI assistance. That is not a binding industry rule, but it reframes training as part of AI governance: efficiency gains can create a long-term capability cost if junior lawyers no longer practice the work they are expected to supervise later.

 

Clients also face a trade-off. Pressure to reduce legal bills with AI can lower drafting time, yet an unreliable filing may require corrections, trigger sanctions proceedings or damage a case. Subramanian’s order leaves the technology available while making clear that accountability remains with counsel.

 

The immediate case ended without sanctions, but the warning is consequential because it addresses both output accuracy and the profession’s talent pipeline. Law firms now have to show that AI controls operate in practice, not merely that policies, training sessions and specialist software exist on paper.