ChatGPT Conversations Become Prosecution Evidence in Australian Court Case
A Sydney man charged over a large-scale data breach of a NSW courts website allegedly used ChatGPT to build the scraping tool that carried out the access, then turned to the same AI for legal advice and coaching on what to tell police. Prosecutors plan to introduce those ChatGPT conversations as evidence in court, a move believed to be an Australian first.
Christopher John Duff, 40, has pleaded not guilty to four counts of accessing restricted data held in a computer. The charges follow a Department of Communities and Justice investigation that found 8,769 restricted documents had been accessed through the department's online registry between January and March 2025.
What the Alleged Scraping Looked Like
The NSW online registry requires a login to access. Police say Duff used ChatGPT to write Python scraper scripts designed to bulk-download documents from the site, circumventing the normal navigation process that limits what a logged-in user can browse one page at a time. Prosecutor Sergeant Andrew Maldigri compared the approach to casting a fishing net into the ocean, hauling it back, and sorting through the catch on deck to find what you were after.
Police say cybercrime detectives investigated the unauthorized access of thousands of restricted documents held by the department. Duff was reportedly helping a friend involved in complicated bankruptcy proceedings at the time.
Defence counsel Daniel McMahon told Judge Lisa Graham at Downing Centre Local Court that there may not be a dispute, in general terms, that his client was responsible for some of the alleged conduct. The core defence argument is about knowledge: McMahon said the offence requires proof that the accused knew at the time that the access was unauthorized, not merely that he was reckless about it. Recklessness, he said, is insufficient.
ChatGPT as Evidence and as a Source of motive
Sergeant Maldigri told the court that prosecutors would rely on ChatGPT conversations and forensic expert testimony, raising an issue he believed had not yet been addressed in Australian courts: generative AI records being used by the prosecution to establish motive and as evidence of the offending itself.
He pointed to US precedents, including a New York fraud case where a defendant unsuccessfully objected to the use of records from AI assistant Claude in the proceedings.
According to prosecutors, Duff turned to ChatGPT for legal advice after his login to the registry had been shut down and before he was charged. He allegedly asked the chatbot what police would do in the situation. He also allegedly sought coaching on what to say if approached by investigators. He ultimately refused to answer questions in a police interview.
McMahon pushed back on the framing. He told the court that the tenor of the ChatGPT queries, if they were made by his client, was more suggestive of a post-fact realisation. The questions reportedly included asking whether authorities knew if "it was even a crime yet?" and whether what had been done was unauthorized.
A Case Buried in Its Own Evidence
The three-day hearing stalled before it could get started. McMahon estimated there were more than 10,000 pages of proposed exhibit material. Lawyers expressed doubt the hearing could be completed even with a full week allocated. The case was adjourned for a fortnight, with a new, longer hearing to be scheduled.
Duff was granted bail in April 2025 on conditions that included surrendering his passport and possessing only one "dumb phone" with call or text functions. The restrictions suggest the court considers the case serious enough to warrant travel and communication limitations while it proceeds.
What This Means for AI Evidence in Court
The Duff case sits at the intersection of two trends that are accelerating fast. The first is the use of generative AI to write functional code. Scraper scripts are not new, but the barrier to building one has dropped sharply. A person with no programming background can ask a chatbot to write a Python script that automates HTTP requests, parses HTML, and downloads files in bulk. The prosecution's theory in this case rests on that exact workflow.
The second trend is the evidentiary status of AI chat logs. ChatGPT conversations are stored by default. They can be subpoenaed. When a defendant turns to an AI for legal advice or coaching, those transcripts become a record of state of mind, which is precisely what the prosecution needs to prove in a case like this. The defence will argue the queries show uncertainty and confusion, not knowledge of wrongdoing. The prosecution will argue they show awareness and intent.
Australian courts have not yet tested this territory. US courts have. The New York case cited by prosecutors involved a fraud defendant whose Claude conversations were admitted as evidence. If the Duff case proceeds to full trial, it will set the first significant precedent in Australia for how AI-generated conversations are treated in criminal proceedings.
For developers and anyone who uses AI tools daily, the case is a reminder that chatbot conversations are not ephemeral. They are stored, they can be retrieved, and they can end up as exhibits in a courtroom. The same tool that helps you write a script can also produce a transcript that prosecutors read very differently than you intended.