It is like a quarterback calling his plays into a megaphone while the opposing defense is standing three feet away. You might think you are strategizing in private, but you are actually just handing over the playbook. That is essentially what happened in the trial of Jonathan Rinderknecht, who is accused of starting the Palisades fire on New Year’s Day 2025. While the prosecution had the usual suspects—iPhone location data, CCTV, and witnesses—they decided to add a new flavor to the evidence pile: ChatGPT logs.
Yes, and it is surprisingly straightforward. As detailed in this report from The Verge, the state doesn’t need to hack your account or find a forgotten password. They simply serve a subpoena to OpenAI. Once the legal paperwork is processed, the company hands over the conversation history. The friction here isn’t technical; it’s just a matter of how fast a corporate legal team can move.
Who actually reads the TOS before clicking agree? We all just skip to the bottom, ignoring the part where we essentially agree that our “private” thoughts are stored on a server in a warehouse we’ll never visit. The logs aren’t just a mirror of the user; they are a permanent, searchable record held by a third party.
This is where the legal system is about to hit a wall. The prosecution is treating these logs as a digital diary or a confession, but that is a fundamentally flawed way to view how people actually use LLMs. Anyone who has spent more than ten minutes with a prompt knows that we treat these models as sounding boards, roleplay partners, or “what if” machines. Asking a model how to do something—or simulating a scenario—is not the same as documenting a plan.
If I ask an LLM to help me write a fictional scene about a heist, am I suddenly a criminal mastermind in the eyes of the law? (Probably not, but a jury might disagree). Treating a prompt as a statement of intent ignores the iterative, often nonsensical nature of human-AI interaction. The prompt is not the person.
Hardly. Deleting a conversation from your sidebar is a cosmetic change, not a forensic one. There is a massive gap between the user interface and the actual database. Most cloud-based AI providers keep logs for safety and training purposes long after the user hits the trash icon.
We have seen this pattern before with early SMS and email archives—the belief that “deleted” means “gone” is a relic of the 90s. If the data is stored on a server, it exists until the company’s retention policy says otherwise. Relying on a “Clear Chat” button to avoid a subpoena is like trying to hide a house by closing the curtains.
Not for the average user, but it will accelerate the migration of the “power user” to local hardware. The realization that your internal monologue is now a discoverable legal asset is a powerful motivator to stop relying on an API. We are moving toward a world where the only truly private thought is one that stays in VRAM.
Within six months, we will see the first major legal precedent that explicitly separates “exploratory prompting” from “criminal intent” in a US court. Until then, the risk is simply too high to trust a cloud provider with anything you wouldn’t want read aloud in a courtroom. If you aren’t running your models locally, you aren’t having a conversation; you’re creating a transcript.
