How ChatGPT stores your data
What OpenAI does with your ChatGPT conversations, how long they are retained, when they are used for training, and how the enterprise and API offerings differ from the consumer product.Retention by default
On the consumer ChatGPT product (Free, Plus, Team, and similar tiers), conversations are retained by default so the model can reference earlier context and so OpenAI can improve the service. You can turn off chat history in settings, which also disables use of your chats for model training. The official details live in the OpenAI Data Controls FAQ, which is the source of truth because OpenAI updates it when policies shift.Training data opt-out
If chat history is on, OpenAI may use your prompts and the model's responses to improve future models. Turning off history disables training use for your account. On the Enterprise and Business tiers, training on your data is disabled by default, which is one of the core differences between the consumer and the enterprise product.For API usage (developers calling OpenAI's endpoints), OpenAI states it does not train on API inputs and outputs by default. This has been the policy since the 2023 API update. Refer to the OpenAI enterprise privacy page for the authoritative statement.Retention for safety review
Even with training opted out, OpenAI retains some data for a limited period for abuse monitoring and compliance. The specifics vary by product, but the general pattern is: short-term retention for safety, regardless of training preferences. For regulated industries (health, legal, finance), this nuance matters, because "we don't train on it" is not the same as "we don't keep it."What this means for your writing
If you draft with ChatGPT and paste proprietary material, three concerns apply:- Confidentiality. Don't paste material you're legally prohibited from sharing (NDAs, PHI, PII, classified content). Even on enterprise tiers, human-review access exists for abuse monitoring.
- Future training exposure. If history is on (consumer tier), your content may shape future models. Sensitive drafts, unique research, and original voice can leak into the training corpus.
- Leakage via other users. There is no documented case of verbatim user content appearing in another user's output, but models are probabilistic, so unusual phrasings seen repeatedly in training data become more likely to appear. For truly sensitive text, the safe answer is "don't paste."
Detection implications
This matters for AI detection because some tools use OpenAI's API directly and subject your text to OpenAI's retention policies. If you're running detection on private material (a student paper, an internal memo), pick a detector that processes locally or documents its own retention. It shouldn't silently forward your text to a third-party LLM provider.Leap's AI detector runs entirely in your browser: free, no account, and nothing is sent anywhere. Your text never leaves your device. Our guide to how detectors work covers the architecture differences.Privacy-first practices for AI-assisted writing
- Turn off chat history on consumer ChatGPT if you're pasting work you want to keep out of training.
- Use the API or Enterprise tier for client work, legal documents, or research, as both have stronger default privacy.
- Strip identifiers. Don't paste names, client names, employer names, or other PII if you can avoid it. Generalize the content before prompting.
- Check third-party tool policies. If you use a browser extension, a writing app, or a humanizer, read its privacy policy. Many forward your text to OpenAI, Anthropic, or other providers.
- Run detection locally where possible. For sensitive documents, prefer a detector that doesn't upload the full text to an external model. Leap's detector runs in the browser, so your text stays on your device.