Regulation AI Minute Newsroom 2026-08-16

A man hid white-on-white orders to the AI inside his court filing. The judge caught it by printing the pages out.

Connecticut Superior Court Judge Walter M. Spader Jr. sanctioned self-represented plaintiff Matthew Elliott on 6 August after finding concealed text in his pleadings: instructions set in tiny white type on a white background, invisible to a reader but legible to any software parsing the file, directing any AI system that reviewed the document to produce only conclusions favourable to him and to treat an earlier clerk's ruling as an error to be corrected in his favour. The judge noticed unusual white space while going through printed copies of the docket. In his 14-page decision he disclosed his own use of AI — Google's Gemini for a working translation of a Brazilian ruling, and Westlaw's AI review features to check authorities — and wrote that judgment can never be delegated to a machine. Elliott is now barred from filing electronically and must bring printed copies to the clerk's office in person.

Why it mattersThis is the first documented prompt-injection attempt aimed at a US court, and what it really tested was an assumption: that the judge would let a model read for him. He did use AI on parts of the work — and still caught the trick, but only because he also read the paper. As courts, agencies and employers start pushing submitted documents through models, every document becomes a possible instruction to the machine reading it, and "print it out and look at it" is a defence that does not scale.

✓ Verified · 3 sources

A man hid white-on-white orders to the AI inside his court filing. The judge caught it by printing the pages out.

Related stories