Does ChatGPT Train on Your Conversations? What Actually Happens to Your Data

The short answer, with the important caveats

On personal ChatGPT accounts, model training is on by default. OpenAI states it may use the content of your conversations to improve the models that power ChatGPT unless you opt out. This applies to Free, Plus, and Pro consumer accounts.

Two things change that default:

• Data Controls opt-out. In ChatGPT, go to your profile, then Settings, then Data Controls, and switch off "Improve the model for everyone." After that, OpenAI says new conversations will not be used to train its models. The setting applies across your whole account.
• Temporary Chat. Chats started in Temporary Chat do not appear in your history, do not create or use memory, and are not used for training. As of 2026, OpenAI says Temporary Chats are deleted within 30 days unless retention is required for safety or legal reasons.

Note the word new. Opting out is forward-looking. Content that was already incorporated into a completed training run cannot be pulled back out of a model. Turning the toggle off stops future use; it does not un-train the past.

What "improve our models" actually means

When a provider says it uses conversations to improve models, it generally means your prompts and the model's responses may become training examples, sometimes after de-identification, and may also be reviewed by humans or automated systems to evaluate quality, safety, and abuse. This is different from a few other things people often conflate:

• Retention is how long your chats are stored, regardless of whether they train anything.
• Human review happens on many platforms for safety and abuse monitoring even when you have opted out of training.
• Memory is a separate feature that lets the assistant remember facts across chats; it is not the same as model training.

So opting out of training does not mean nothing is stored or seen. Most providers keep logs for safety, legal compliance, and abuse detection independent of your training preference.

The retention wrinkle: deleted does not always mean gone

There is a live legal complication worth knowing about. In the copyright litigation between The New York Times and OpenAI, a court ordered OpenAI in May 2025 to preserve ChatGPT output logs, including chats users had deleted that would otherwise be removed within roughly 30 days. That broad, going-forward preservation order was scaled back in October 2025, so deleted chats and Temporary Chats again purge within about 30 days for most people. However, logs already captured under the earlier order, and data tied to accounts the plaintiffs specifically flag, must still be retained.

Separately, in November 2025 a magistrate judge ordered OpenAI to produce 20 million de-identified chat logs to the news plaintiffs, and in January 2026 a district judge affirmed that order over OpenAI's objection that it was a privacy overreach.

The practical takeaway: a legal hold can override a provider's normal delete-and-purge timeline. This is not unique to OpenAI, but it is a concrete reminder that "delete" on a large consumer platform can mean "hidden from you and scheduled for deletion," not "already erased everywhere" — and that litigation can freeze that timeline.

Where API, Enterprise, and Team differ

OpenAI states that, by default, it does not train on inputs or outputs from its business products, including ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, and the API. If you build on the API or your company uses Enterprise, training is off by default rather than opt-out. This is a real and meaningful difference from the consumer app, and it is why many privacy-conscious users route sensitive work through the API or a business plan.

How other assistants compare

The default matters, and it varies by provider:

• Claude (Anthropic). Under a policy change that took effect in 2025, consumer Claude conversations (Free, Pro, and Max) can be used for training unless you opt out, and choosing to allow training extends data retention to as long as five years. Anthropic's policy also notes that conversations flagged for safety review can still be used regardless of your preference. Business and API usage are governed by separate commercial terms. Check your current settings rather than assuming.
• Gemini and Grok similarly lean toward using consumer activity by default, with opt-out controls; details are version-dependent, so verify in-app.
• Duck.ai (DuckDuckGo). Requests are anonymized with identifying metadata like your IP stripped before they reach the model provider. DuckDuckGo says it does not use chats for AI training, and it has agreements with the underlying providers barring them from using prompts and outputs to train, with deletion within 30 days. Recent chats are stored locally on your device by default rather than on DuckDuckGo's servers.
• Proton Lumo. Uses zero-access encryption and states it keeps no logs and never trains on conversations, and it is open source.

Where Evenfall fits

Evenfall is one option among these, and it is worth being precise about what it does and does not do.

What it does: it does not train on your conversations, ever. It runs no ads and no behavioural tracking, ad pixels, or cross-site profiling. When you delete a conversation, it removes the database row for real, with no soft-delete, shadow archive, or recovery window, and account deletion is immediate. You can sign in with an email one-time code or Google, and pay anonymously with a one-time 30-day crypto pass (no card, no name on the invoice) if you prefer. Card payments go through Stripe and never touch Evenfall. It is a fast, full-history chat app with streaming responses, folders, custom instructions, automatic web search, and voice mode on Pro, using a frontier model chosen from a picker.

Evenfall's privacy is enforced by policy — no training, no ads, no tracking, and real deletion — with data encrypted in transit and at rest, and it keeps your history synced across devices until you remove it. (There is no public API, native mobile app, or file upload yet.)

If your priority is "my chats are never used to train a model and delete means delete," Evenfall is built for that. If your priority is "the provider must be mathematically unable to read my chats," a zero-access option like Lumo is the stronger fit. See our principles or compare options to decide what matters most for you.