Courts meet AI text as evidence, covert prompt and stripped mark
OpenAI handed ChatGPT logs that carried a Florida stalking case into court. In Connecticut a plaintiff who hid invisible AI instructions lost electronic filing rights, and a free tool now strips Claude attribution marks.
Artificial Intelligence··Night
Chat logs enter a Florida stalking case
OpenAI reported a 25-year-old Palm Beach analyst to the FBI after he set out plans in ChatGPT to rape and murder his former girlfriend. Federal agents passed two months of his chat logs to the Palm Beach County Sheriff's Office in May, and deputies arrested him that month. A deputy who read the logs said they showed a pattern of rehearsal and planning rather than vague emotional outbursts. The file handed over held only the man's messages, not the chatbot's replies. He was charged in June with aggravated stalking, written threats to kill and illegal use of a cell phone, felonies carrying up to 25 years. He pleaded guilty on 13 August under an agreement his former girlfriend approved, and Circuit Judge Scott Suskauer withheld adjudication and imposed 8 years of probation with two years of electronic monitoring. The case turns private model prompts into court-ready evidence once a provider decides the content meets a reporting threshold.[1]
Hidden prompts cost electronic filing rights
In Connecticut a different use of model language met the court as procedural abuse. In Matthew Elliott v. New York Bariatric Group, the plaintiff placed prompts hidden in 3-point white text on a white background; the court noticed the unusual whitespace. The case, filed in October 2025, alleged data privacy violations and discrimination. Judge Walter Spader Jr. scheduled a hearing and explicitly warned Elliott against hiding text. Elliott ignored the warning and kept hiding messages, calling the first attempt an audit and the later ones invisible jokes. Spader pointed out how plainly improper it would be for a party to arrange for an automated agent to communicate covertly with a juror. Because Connecticut courts do not use AI systems for review, the hidden instructions had no effect on the outcome. The court still stripped Elliott of his right to file electronically and required paper submissions in person. Here the model arrives as a channel the litigant tried to open without the other side seeing.[2]
A free tool strips Claude attribution marks
Outside the courtroom, attribution itself is under pressure. Cardano founder Charles Hoskinson published a free command-line tool called Anthropies on GitHub that targets three attribution markers on Claude output: the keyed text watermark, C2PA credentials signed into image files, and Co-Authored-By trailers left in Git commits. The tool offers a cleaning function that removes banners and commit trailers, plus a rewriting function that restructures prose through models other than Claude so the watermark is not simply reproduced. Hoskinson framed the project around ownership, attribution and digital provenance, and described it as a skill usable with most large language models. Anthropic had switched the marking on across Claude products days earlier; the company has published no response to the tool. Read with the Florida logs and the Connecticut white-text prompts, the three developments show AI text entering institutions in opposite directions: as evidence a provider hands to police, as a covert instruction a party tries to smuggle into a docket, and as provenance marks a public tool is built to erase. Courts and users are already treating model text as something that can convict, mislead, or be scrubbed of origin.[3], [1], [2]
