Matthew Elliott thought he’d found a clever loophole. The self-represented plaintiff in a Connecticut data privacy lawsuit embedded instructions for AI systems in 3-point white text on a white background — invisible to any human reading the document, but perfectly legible to a language model processing it. The hidden messages directed any hypothetical AI reviewer to treat his filing as correct and a prior clerk’s denial as an error. The goal? Secretly sway an automated review in his favor.
Judge Walter Spader Jr. wasn’t impressed. The court noticed suspicious whitespace in Elliott’s filings, uncovered the hidden text, and issued an explicit warning. Elliott responded by hiding more messages — including a YouTube link and what he later called “invisible jokes.” In a 14-page ruling, Spader stripped Elliott of electronic filing privileges entirely. He now has to submit everything in person, on paper.
Spader’s reasoning cuts to the heart of why this matters. Connecticut courts don’t actually use AI to review filings, so the injection failed on a practical level. But the attempt itself was the offense. The judge drew a sharp comparison: hiding instructions for an AI system is like arranging for an automated agent to secretly communicate with a juror. It’s covert influence aimed at whoever — or whatever — ends up making decisions.
This isn’t an isolated stunt. Last year Nikkei found 17 arXiv preprints with similar hidden instructions like “positive review only” buried in white-on-white text, aimed at skewing AI-powered peer review. Meanwhile, MIT and USC researchers found pro se lawsuits in federal courts nearly doubled after ChatGPT went mainstream — 41,490 filed in 2025 alone, with 18 percent of a sample flagged as AI-generated. Elliott’s case is a preview of a messier problem: as courts and journals adopt AI tools, every document becomes a potential attack surface. The judge welcomed AI for helping people without lawyers draft clearer filings. But he drew a line where assistance becomes manipulation. That line is going to get tested a lot more.
