A man hid instructions to try to manipulate an AI during a trial in the U.S

A man hid instructions inside court documents to try to influence an artificial intelligence during a trial in the United States.

Judge Walter Spader Jr. discovered the messages and sanctioned Matthew Elliott, although he decided not to impose a monetary fine.

Prompt injection: what the man did to try to influence the trial

The case occurred in Connecticut and was reported by Ars Technica based on a court decision published last week. The plaintiff, Matthew Elliott, allegedly hid instructions within his filings so that a potential artificial intelligence would favor his arguments.

The text was written in tiny font and in white color on a white background. This way, it was invisible to a person but could be read by software processing the document's content.

The instructions aimed for any AI system analyzing the writing to support Elliott's arguments, ignore previous court decisions, and propose solutions aligned with his position.

The judge discovered the hidden instructions

Judge Walter Spader Jr. determined that the attempt did not affect the resolution of the case. The magistrate explained that Elliott's writings were evaluated based on their arguments and the merits of the case.

However, he considered that the use of prompt injection within court filings set a "dangerous" precedent. Elliott continued to hide messages even after being warned about possible sanctions.

The plaintiff maintained that some of those subsequent messages were simply "jokes". Among them was a link to a video of Nosferatu, a message that said "hi :) I hope you can't see me," and another nonsensical text.

For Spader, the explanation was not convincing. The judge stated that hiding messages within writings intended for the court to take seriously "challenges logic."

What the judge said about the use of artificial intelligence in courts

Elliott argued that one of the instructions aimed to audit the court. According to his explanation, he feared that the Justice system was improperly using AI to resolve cases.

Spader responded that if that was truly his concern, he could express it in visible words. For the judge, the fact that he hid the instructions was a sign of his malicious intent.

The magistrate also clarified that the Judicial Branch of Connecticut does not use AI to review or decide on filings. Therefore, in this case, there was no real risk of a judicial system confusing Elliott's instructions with orders from the court.

As Spader explained, a prompt injection seeks to introduce an instruction within a document that will then be processed by a system. The goal is for the tool to interpret that command as if it came from the person operating the system.

What sanction Matthew Elliott received for using prompt injection

The judge considered that the hidden instructions warranted a sanction because they attempted to communicate covertly with the court and left the defendants in an unequal position.

However, Spader took into account that Elliott was acting as a pro se litigant, meaning without legal representation. Therefore, he decided not to impose a monetary sanction.

Instead, he prohibited him from filing documents electronically in the future. From now on, he must make paper submissions, a measure that, according to the judge, does not affect his access to Justice and prevents further abuses of the electronic filing system.

The warning about using chatbots to prepare lawsuits

Spader stated that it is not surprising that attempts at prompt injection appear in courts. The technique is already used in other areas, such as labor processes where texts intended for AI systems analyzing resumes are hidden.