Prosecutors in the Los Angeles Palisades wildfire arson trial built their case on iPhone location data, security footage, and witness testimony — and then, in a national first, on the defendant’s own ChatGPT logs. The jury wasn’t convinced: it deadlocked 10-2 in favour of acquittal, triggering a mistrial and putting a hard asterisk on the idea that AI chat history can stand as proof of criminal intent.
THE BOTTOM LINE
Your private prompts to a commercial LLM are now prosecutable evidence. That fact survives the mistrial, and it should worry every Kiwi using ChatGPT, Claude, or Gemini for anything more personal than a recipe. The Palisades case is the first time a US court has put generative-AI conversation logs in front of a jury as proof of state of mind, and the result — 10 jurors unconvinced by material the prosecution treated as central — tells us less about this defendant than about the evidentiary ceiling of chatbot history 28 Million Secrets, 47,000 Backdoored Machines: The AI Agent Security Crisis Nobody Solved.
The Palisades fire and the digital trail prosecutors followed
The charges stemmed from the New Year’s Day 2025 Palisades fire in Los Angeles, one of the most destructive wildfires in California history. According to The Verge’s reporting on the mistrial, the prosecution’s conventional case was already substantial: cell-site location data placed the defendant near the ignition point, security cameras captured relevant movements, and civilian witnesses placed him in the area during the critical window.
The novel layer was the ChatGPT history. Prosecutors introduced conversation logs and image-generation prompts they argued demonstrated premeditation and a grievance-rich state of mind — including prompts about fire imagery, a query along the lines of “why am I so angry all the time,” and broader venting about wealthy individuals allegedly “destroying the world.” They also entered a screen recording in which the defendant appeared to muse about responsibility for a fire started by a cigarette.
On paper, this looks damning. In a jury room, it landed very differently.
Why a 10-2 deadlock on the defence side actually undermines the prosecution’s theory
The verdict shape matters more than the mistrial headline. A 10-2 split favouring acquittal is not a near-miss for the prosecution — it is a near-verdict that the chat logs, even stacked on top of everything else, did not prove intent beyond a reasonable doubt.
The most telling post-trial voice came from a juror who spoke to CBS Los Angeles after the mistrial was declared. She said she uses ChatGPT herself, regularly, and that the prosecution’s chat-log evidence struck her as nothing more than a “character flaw” — venting that anyone might type into a chatbot at 2am without ever acting on it. That framing — character, not conduct — is exactly the distinction criminal intent requires, and it is precisely the distinction the prosecution’s AI evidence failed to draw.
This is the uncomfortable part for law enforcement: the same affordance that makes ChatGPT useful (it will engage with dark ideation in text form) makes its outputs useless as a confession. A model that will write you a manifesto will also write you a rant you have no intention of carrying out. Juries appear to grasp this intuitively; prosecutors evidently did not get the memo.
The privacy precedent is bigger than the verdict
The mistrial does not erase the discovery process that produced the logs. To get those conversations in front of the jury, prosecutors had to compel OpenAI — or obtain them through a cooperating defendant’s account — and introduce them over authentication and relevance objections. That mechanism now exists, has been tested under oath, and will be reused.
OpenAI’s privacy policy and terms of service permit disclosure in response to valid legal process, which is standard across the industry and not a surprise. What is new is the demonstrated appetite: this was an arson-murder mass-casualty case, but the legal pathway it validated — ChatGPT logs as evidence of intent in any serious criminal matter — will not stay confined to mass-casualty prosecutions. Expect district attorneys to file the same motions in fraud, harassment, stalking, and threat cases over the next twelve months.
What this means for New Zealand users under the Privacy Act 2020
Kiwis are not insulated from this trajectory just because the trial was in LA. OpenAI, Google, and Anthropic all host data outside New Zealand, but the Privacy Act 2020 still binds how local agencies, employers, and courts handle information about you — and any AI logs they hold or subpoena.
Three practical implications:
- Information privacy principle 6 requires that personal information be collected only for a lawful, connected purpose. A venting prompt typed at 11pm is not consent to future evidence in a criminal file. The US outcome shows the gap between technical policy and courtroom reality.
- Cross-border discovery is asymmetric. US prosecutors can compel US-headquartered AI firms; New Zealand Police generally cannot compel OpenAI directly and must work through mutual legal assistance. That gap will narrow as AI firms establish local data-residency arrangements and as New Zealand’s law-enforcement access frameworks mature.
- Employers using AI tools for HR, recruitment, or monitoring sit on a related risk surface. If a worker’s ChatGPT or Copilot history can be subpoenaed in a US case, the same logs can be requested under the Discovery Rules in New Zealand civil and employment proceedings. The conversation you have with a workplace chatbot has no special protection just because it feels casual.
The deeper question is whether commercial LLMs should carry the same expectation of confidentiality as a letter to a lawyer or a doctor’s notes. Right now, they do not — and most users, here and overseas, do not know that.
FAQ: What every AI user should take away from the Palisades mistrial
Can my private ChatGPT conversations be used against me in court? Yes. As the Palisades case shows, they can be subpoenaed from OpenAI and presented as evidence of state of mind, motive, or intent. The mistrial does not undo that pathway — it only limits how much weight a jury will give the resulting material.
Does an angry or violent prompt to an LLM prove I committed a crime? No. The 10-2 deadlock suggests juries treat chatbot output as expression, not action. A prompt is not a plan, and a rant is not an admission. That said, prosecutors will continue to introduce them, and the cost of defending against AI-derived evidence at trial is real even when the defence wins.
What should I stop putting into commercial AI chatbots? Treat any commercial LLM like an unencrypted email to a stranger — not like a private journal. Stop typing confessions, threats, specific targeting of people, or anything you would not want read aloud in a courtroom. Use local or on-device models for genuinely sensitive thinking until legal protections catch up.
Does New Zealand’s Privacy Act 2020 protect my AI chat history? Partially. The Act governs how New Zealand agencies handle your personal information and gives you rights of access and correction, but it does not override a US subpoena to a US-headquartered AI provider, and it offers limited protection once data is held offshore. The Act is a useful tool against local misuse, not a shield against foreign legal process.
Will this case change how prosecutors use AI evidence going forward? Yes, but probably not the way the headline suggests. Prosecutors will keep filing the motions — the mistrial cost the prosecution nothing material, and the next case will have stronger corroborating evidence. The shift will be tactical: chat logs as a framing device around a conventional case, not as a standalone proof of intent.
THE BOTTOM LINE
The Palisades mistrial is not a defeat for AI evidence — it is the first successful deployment of it. Prosecutors have now shown they can put your chatbot history in front of a jury; the Palisades jury simply declined to convict on it. That is a ceiling, not a wall, and every Kiwi using a commercial LLM for anything sensitive should be planning around it. Until legal frameworks treat LLM conversations with the same default confidentiality we apply to lawyers, doctors, and journalists, the only reliable safeguard is what you choose not to type.
