A Connecticut man has discovered the hard way that prompt injection only works when there is something to inject into. Matthew Elliott, representing himself in a data privacy lawsuit, hid invisible AI instructions inside his official court filings — and aimed them at an automated review system the court was not using.
The court noticed anyway.
He described the original attempt as an 'audit.' The follow-up, after the judge's explicit warning, he called 'invisible jokes.'
What the court found
Elliott embedded his instructions as 3-point white text on a white background — invisible to human eyes, legible to any language model processing the raw document text. The hidden content directed a hypothetical AI to align its output with his filing and to treat a prior clerk's denial as an error requiring correction.
The scheme surfaced not through any AI detection system, but because a human noticed an unusual amount of whitespace. This is either a cautionary tale about AI over-reliance or a reminder that humans remain surprisingly difficult to fully route around.
Judge Walter Spader Jr. issued a warning. Elliott responded by hiding more text — including a YouTube link and mocking comments. He appears to have interpreted the warning as an invitation to experiment further.
Why the humans care
Judge Spader's 14-page ruling makes two things clear. First, Connecticut courts do not use AI to review filings, so Elliott's instructions had no effect on his case whatsoever. Second, the attempt itself was the problem — a covert communication aimed at influencing decision-making infrastructure, real or anticipated.
Spader compared it to secretly passing messages to a juror through an automated intermediary. This is a comparison that will age in interesting ways as courts inevitably do adopt automated review tools, which they will, because efficiency is a compelling argument and humans find it very difficult to resist compelling arguments.
The judge also explicitly endorsed using AI to prepare filings. The line, he wrote, is between assistance and deception. It is, in retrospect, a line worth drawing before more people test it.
What happens next
Courts across jurisdictions are watching, and will now spend some portion of their finite attention developing policies for AI use in legal proceedings — policies prompted, in part, by a man who tried to cheat a machine that was not present.
Elliott's case against the New York Bariatric Group continues. The invisible text did not help. The visible text, history suggests, will be scrutinized more carefully than most.