How AI training data works
At a glance
- Training adjusts a model’s parameters using selected data. Public availability alone does not establish that your data was selected.
- A service can also retrieve web pages or use saved chats, files and memories when producing an answer. These are different from training.
- Training preferences, stored-content deletion and requests about a response have different scopes. Compare provider controls.
What training changes
A model’s parameters are numbers adjusted during learning. Language-model training can use prediction tasks to learn patterns; later training can adapt capabilities or behavior. Data selection and training methods differ by model.
If personal data is included, training may affect the parameters. That does not establish that a particular person’s data was included or that a response faithfully reproduces it. A privacy request needs evidence about the specific issue.
Where training data can come from
OpenAI describes public internet information, partner data and material supplied or generated by users, trainers and researchers, including synthetic data. This is one provider’s description, not an inventory for every model. Its documentation also describes filtering to reduce personal data, including excluding known sources that aggregate it.
Common Crawl provides an open corpus of web pages, metadata and text extracts. Finding a page in a crawl establishes a crawl record, not that a particular model selected it for training.
Provider documentation may describe source categories without identifying each individual record. Avoid treating a public listing, a crawl record or a plausible AI answer as a complete training history.
Training, retrieval and saved content are different
- Training: data is used to adjust model parameters.
- Retrieval: a service supplies information when answering, such as a current web page. Microsoft documents web search as a separate Copilot control.
- Stored content: chats, uploads, memories or connected services may provide context. Each has its own controls and retention rules.
An answer can combine these sources or contain an error. If it cites a public page, check that page directly and preserve the relevant URL. Removing it and reporting the AI response are separate actions.
What each kind of evidence shows
Each kind of evidence answers a narrower question than whether your personal data was used for training. The table explains what each one shows; this is general explanation, not a statement about any particular model’s training data. For the steps, use the general AI checklist.
| Evidence | What it shows | What it can’t show | Next step |
|---|---|---|---|
| A public listing | The page showed your personal data when you checked it | That any model was trained on it | Ask the site to remove it |
| A crawl record | A crawler such as Common Crawl captured the page at a recorded time | That a particular model selected that capture for training | Keep the record and address the live page |
| A response that cites a page | The response linked that page when it was produced | That the page was in training data, or that the response matches the page | Check the cited page, then report the response if it’s still a problem |
| A response with no source | The service produced that text at that time | Where the text came from: training, retrieval, saved content or an error | Save the prompt, response, date and product, then report the response |
| A chat in your own account | The provider stores it as account content under its policy | Whether it was used for training, which depends on the product, your settings at the time and exceptions such as feedback | Delete saved content or stop future training |
What the privacy controls cover
OpenAI offers a choice for new eligible conversations; feedback can still include the associated conversation for training. Anthropic’s consumer choice has separate feedback, safety-review and explicit opt-in rules. A paid subscription alone does not establish training permission.
Deleting stored content follows the service’s policy and exceptions. Neither a deletion confirmation nor a training preference alone proves a change to previously completed training. A response request also needs its own confirmation and follow-up.
There is no single provider setting that removes personal data from every service and public source. Use the general AI privacy checklist, then follow the documentation for your product, account and region.
What you can do now
- Review your account’s policy. Check the product, plan, region and saved training preference before sharing personal data. Read the rules for feedback and temporary or incognito use.
- Manage stored content separately. Check chats, files, memories and connected services. Keep any deletion confirmation and stated exceptions.
- Document a concerning answer. Save the prompt, response, date, product and cited URLs. Use the provider’s privacy or personal-data channel when available.
- Address public sources. Review exposed broker listings and other pages. Removal can reduce exposure there; it does not prove that a model changed.
- Record the outcome. Keep the provider’s reply and recheck the specific output or source. Do not equate a submitted request with removal.
Delist finds exposed personal data, pursues supported removal requests and tracks their status. The free scan is not an inspection of private training datasets.
Run a free scan →Frequently asked questions
Does deleting a chat make a model forget it?
Chat deletion manages stored content under the provider’s retention policy. It does not, by itself, establish that previously completed model training has changed. Training preferences and response-removal requests are separate actions.
Was my data definitely in ChatGPT or Claude’s training?
A public web presence or an answer about you does not prove that a particular model used your data in training. Look for evidence about the specific dataset or service, and check whether the answer cites a current source.
Can I tell whether an answer came from training or search?
A source link can help you investigate retrieved content. Without product-specific evidence, the answer alone does not establish where a fact came from. It may also contain an error.
Do AI providers train on my conversations?
It depends on the service, account and saved preference. OpenAI and Anthropic document consumer controls and separate feedback rules. Safety processing and commercial product terms can differ. Review the current policy for the account you use.
Sources
- OpenAI: model development and personal-data requests
- Common Crawl: web-corpus overview
- OpenAI: model-improvement controls and feedback
- Anthropic: consumer training choice and exceptions
- Microsoft: updated Copilot privacy controls
Sources accessed . See our source standards.