Key Takeaways

  • A Connecticut Superior Court judge revoked a self-represented plaintiff’s e-filing privileges after he repeatedly embedded invisible “white-on-white” text in court submissions directing any AI system reviewing the filings to rule in his favor. 
  • Courts may treat concealed AI-directed instructions as an independent form of litigation abuse, sanctionable even when no AI tool is actually used or deceived. As judicial AI adoption expands, litigants should also be alert to the risk that hidden instructions embedded in an opponent’s filing or a client’s own documents could silently influence AI-assisted review.

In Elliott v. New York Bariatric Group, LLC, No. AAN-CV-25-6066141-S (Conn. Super. Ct. Aug. 6, 2026), an action concerning access to health care records, the court found that the plaintiff’s motion contained hidden, tiny-point white text directing any AI model reviewing the filing to generate output favorable to the plaintiff. Mem. of Decision at 1. The plaintiff repeated the tactic in multiple filings and continued after an explicit warning.  The tactic failed, as the Connecticut Judicial Branch does not use AI to review or decide filings, and the presiding judge denied the underlying motion after reading a printed copy. Id. at 8.  Nonetheless, the court imposed sanctions. Id. at 2.

The Court characterized the conduct as “prompt injection”: an attempt to smuggle a hidden command into the text stream that an AI system processes. Id. at 6. Drawing an analogy to improper ex parte communications, the court reasoned that prompt injection seeks to influence the decisional process through a channel the opposing party cannot see or answer. Id. at 7. Critically, the court held that the attempt itself was sanctionable regardless of whether it succeeded. A concealed falsehood remains improper “even when the person it was meant to deceive happens never to read it.” Id. at 8.

Finding no U.S. authority directly on point, the court cited a May 2026 decision from Brazil’s Eighth Regional Labor Court sanctioning attorneys who embedded similar white-on-white text to manipulate the court’s AI docketing tool. Mem. of Decision at 9–10. While Connecticut did not use AI, the court deemed that distinction immaterial because the concealed instructions in Elliott were “aimed at whatever tool any reader might use.” Id. at 10.

As a sanction, the Connecticut court rescinded the plaintiff’s e-filing privileges, requiring future filings to be made in person on paper. Id. at 13–14. Notably, the court emphasized the importance of integrating the use of AI into legal practice and clarified that litigants remain free to use generative AI in preparing filings so long as they independently verify the output, consistent with Connecticut practice. Id. at 14.