US Government Moves to Bar Defense from Using Public AI Tools With Sensitive Evidence
Bot Mutiny |
Federal prosecutors in Oklahoma filed a motion to prevent the defense team in United States v. Simmers from uploading grand jury materials and medical records to generative AI models.
Federal prosecutors don't want sensitive discovery materials ending up in the training data of a large language model. On September 17, 2026, the U.S. government filed an unopposed motion for a protective order in the Northern District of Oklahoma. The filing targets how the defense handles evidence in the case of Abraham James Yahv Simmers II. The document, first reported by CourtListener, explicitly addresses the threat of data leakage through artificial intelligence tools. Assistant United States Attorney Aaron M. Jolly is asking the court to set hard boundaries on how the defense team uses automated systems to process evidence. The government isn't just worried about typical leaks. They're worried about the black box of model training. The motion seeks to prohibit the defense from inputting or uploading protected materials into any AI tool that retains data for model training. It also seeks a total ban on submitting protected information to any "publicly accessible AI system" that uses data to train its models. ## The Scope of Protected Evidence The evidence in question isn't just bureaucratic paperwork. According to the motion, the discovery includes medical records of an alleged assault victim and personal identifying information of various individuals. It also covers grand jury materials, which are usually kept under tight seal under Federal Rule of Criminal Procedure 6(e). The prosecution argues that these protections are necessary to maintain the "sanctity and the secrecy of the grand jury process." They're trying to prevent a scenario where sensitive witness testimony or private medical history becomes part of a commercial AI's permanent memory. ## AI Restrictions as the New Standard The proposed order allows for AI use only under two strict conditions. First, the tool must not retain or use the materials for training. Second, the tool must not expose the materials to unauthorized third parties. This creates a high bar for defense teams looking to use automated tools for document review or trial preparation. This isn't just a suggestion from the U.S. Attorney's office. The motion states that the defense counsel, Daniel Medlock, does not object to these terms. Both sides seem to agree that the risk of a third-party AI company ingesting federal evidence is too high to ignore. ## Legal Precedents and Requirements The motion cites several foundational legal requirements, including the Jencks Act and the government's obligations under Brady v. Maryland and United States v. Giglio. Usually, the government provides grand jury transcripts voluntarily during discovery to avoid trial delays. But the digital nature of modern legal work has changed the stakes of that disclosure. If the court signs off, it sets a clear boundary. A defense team can't just drop a thousand pages of sensitive grand jury testimony into a public chatbot to summarize it. The government is treating these AI systems as a potential leak to a third party, not just a private tool. The case, docketed as 26-CR-341-JDR, highlights the growing friction between traditional court secrecy and the convenience of generative tech. You can't un-train a model once it eats your data. The DOJ is clearly waking up to that fact.