Your AI chats are discoverable now

A quieter ruling from early 2026 may matter as much to daily practice as the privilege cases. In the consolidated copyright litigation against OpenAI, a federal court affirmed an order compelling production of a twenty-million ChatGPT-log sample. The privacy objection was rejected for a straightforward reason: users had voluntarily handed their conversations to the provider.

Set that beside the privilege ruling in Heppner and the direction of travel is clear. Courts are treating AI prompts and outputs the way they have long treated email, chat, and server logs. They are discoverable. They persist. And a front-end delete button does not save you once a litigation hold attaches, because the record sits on the provider's systems, not yours.

For a legal team the practical rule writes itself. Assume that anything a client routes through a third-party AI is reachable in later litigation, whether the matter is theirs or an adversary's. That is not a reason to avoid these tools. It is a reason to be deliberate about what you put into them.

The most reliable way to keep something out of discovery is to never disclose it in the first place. If the identifying details never reach the tool, there is nothing on the provider's servers to produce. That is the whole idea behind processing locally: replace the sensitive values on your own machine, send only placeholder text to the model, and restore the real values locally when the answer comes back. The document itself stays on your computer.

We built RedactLocal to make that the path of least resistance rather than a discipline you have to remember. The point is not to fear the tools. It is to be intentional about the one moment that matters, the moment data would otherwise leave your machine.