A Connecticut judge has drawn a crisp line for the AI era: using artificial intelligence as an aid in preparing court filings is fine—provided the user checks its work. But hiding instructions designed to manipulate someone else’s AI will not fly.
In Elliott v. New York Bariatric Group, a self-represented plaintiff embedded tiny white-on-white text in court filings. Invisible to ordinary readers but readable by software, it instructed any AI reviewing the documents to agree with the plaintiff and help overturn an earlier ruling.
The plaintiff called it an “audit.” The court called it dishonest.
The plaintiff’s explanation became even harder to swallow when the hidden messages continued after the court identified the practice and warned that sanctions were possible. Later filings included a concealed “hi :)” message and a link to a Nosferatu video.
The prompt never worked. The court does not use AI to decide filings, and the judge denied the motion after reading a printed copy. But the attempt itself was enough: the plaintiff lost access to electronic filing and must now submit papers in person.
Importantly, the court did not reject AI. It welcomed its honest use, particularly as a tool for improving access to justice. The judge even acknowledged using AI-assisted tools while preparing the decision—while independently checking the work and retaining responsibility for the result.
That distinction matters. AI can help draft, summarize, and organize. A hidden prompt is different. The court likened it to a secret communication aimed at the decision-maker—or the tools the decision-maker relies on—without the other side knowing or having a chance to respond.
The lesson is refreshingly simple: use AI if it helps. Check its work. Own the result. But don’t try to make someone else’s machine take your side.
Technology changes. The duty of candor does not.