Where your ChatGPT conversations actually go

Most people type things into ChatGPT they'd never put in an email. Health questions, money problems, half-formed business ideas, code they don't own. So it's worth knowing exactly where that text goes after you hit enter. The short version: to OpenAI's servers, where it's stored, possibly used for training, and (as a 2025 court order demonstrated) sometimes kept even after you delete it.

The default path of a ChatGPT message

You type, it travels to OpenAI's infrastructure, the model answers, and the exchange is stored in your chat history server-side. On consumer plans, conversations can also feed model training unless you find the toggle and switch it off (Settings, Data Controls). That's not a scandal; it's in the documentation. But documentation is read by approximately nobody, so the practical default is: your words become part of the company's data estate.

Enterprise and API tiers carry stronger commitments, including no training on customer data by default. The protections scale with how much you pay, which tells you something about which direction the incentives point.

What the 2025 court order proved

During the New York Times' copyright lawsuit against OpenAI, a federal magistrate ordered the company to preserve ChatGPT output logs as potential evidence, sweeping in conversations users believed they had deleted. OpenAI objected publicly, calling it at odds with its privacy commitments, and the order was later narrowed. But the lesson stands on its own: data held by a third party is subject to that third party's legal exposure, not your intentions. Your delete button lost to someone else's lawsuit.

This isn't a ChatGPT problem. It's a cloud problem

Anthropic moved in 2025 to training on consumer Claude chats unless users opt out, with multi-year retention for those who opt in. Google's Gemini documentation discloses that human reviewers may read sampled conversations. Every cloud assistant works this way to some degree because the architecture demands it: your text has to land on their machines for the model to run. Once it's there, it lives under their policies, their retention schedules, and their subpoenas.

The version where there's nothing to preserve

A local model inverts the architecture. The model runs on your Mac, the conversation is written to your disk, and no third party ever holds a copy. There's no server-side history to train on, no retention policy to read, and nothing a court can order preserved from a provider, because the provider never had it. Outlier works this way: after the one-time model download, you can run it with wifi off and the chat never exists anywhere but your machine.

Honest scope note: local privacy protects you from third-party exposure, not from someone with access to your physical Mac. Disk encryption and a login password are still your job.

Frequently asked questions