Does My AI Provider Use Customer Conversations for Training?
Provider training rules depend on the product and connection. Compare current OpenAI, Anthropic and Gemini terms and distinguish training from retention.
It depends on the provider's specific product, account, connection and settings. Some commercial services exclude customer inputs and outputs from training by default. Consumer products, optional feedback and unpaid developer services can have different rules. "We use this provider" is not enough information to answer the question.
When Niwa connects an assistant to your WordPress business, review the terms for the connection you will actually use. Your website's saved conversation records and the provider's handling of the request are separate layers.
Training, processing and storage are different
Training uses data to develop or improve a model. Processing a prompt to produce an answer is inference. Retention concerns how long content or related records remain stored.
A supplier can exclude your messages from training while still processing them, retaining logs, storing a conversation or reviewing material for safety. A zero-retention arrangement can also be limited to particular accounts, models, endpoints or features. Read the precise scope rather than converting one statement into "the provider never sees or keeps customer data."
This distinction matters for the privacy information you give visitors. Explain the actual purposes and handling rather than promising that every copy stays on your website because training is disabled.
OpenAI: commercial services and individual accounts differ
OpenAI's current model-improvement guidance says business products and the API do not use inputs and outputs for model improvement by default. Individual services have separate controls, and submitting feedback can change the handling of the associated conversation.
Do not apply a ChatGPT Business statement to an individual account simply because both display the ChatGPT name. Likewise, opting out of training does not by itself determine retention. For an API connection, use the API data-control documentation to check the endpoint, state and available account controls.
Anthropic: check commercial use, feedback and model requirements
Anthropic's commercial training guidance excludes commercial chats and coding sessions from training unless the customer participates in relevant opt-in arrangements or supplies material through mechanisms such as feedback. Consumer products have their own policy.
Its commercial retention guidance distinguishes API data from conversations stored in products. It also describes exceptions and additional retention and review requirements for certain Covered Models. Check the selected model and agreement; avoid an unconditional promise that all Anthropic use has identical handling.
Gemini API: paid, unpaid and regional terms matter
The Gemini API Additional Terms, effective 23 March 2026, distinguish paid and unpaid services. Unpaid services can use submitted content and responses for product and machine-learning development, with human review. Google instructs users not to submit sensitive, confidential or personal information to those services.
Paid-service data-use terms exclude prompts and responses from product improvement. The terms also include particular treatment for the EEA, Switzerland and UK. For Gemini API, verify the Cloud project and active billing association, rather than assuming another paid Google account establishes the policy for this request.
Specific features can have additional terms. Also check permitted uses and audience restrictions. The training policy is only one part of deciding whether a connection fits your customer workflow.
Verify the actual route used by your assistant
Write down the provider, account type, API or supported subscription route, selected model and any intermediary. Identify the account that sends the request, not merely the brand selected in a menu.
Then check these points against current terms:
- Whether inputs and outputs are used for model improvement by default.
- Whether the account has enabled a voluntary sharing or training programme.
- Whether feedback, debugging or additional features change handling.
- What retention and review apply to the selected model and feature.
- Which contractual and regional conditions govern the account.
Keep the date and evidence for the decision. Recheck when the model, plan or provider terms change. A screenshot of a setting can support the record, but it does not replace the terms describing what that setting covers.
Choose a connection suited to customer data
Niwa supports eligible own-subscription connections with no added Niwa AI usage fees or Niwa-imposed usage caps on that connection. The underlying provider's terms and limits still apply, and API billing is separate. A cost arrangement does not establish permission to process every type of customer information.
For an ordinary shop, start by answering public catalogue and delivery questions without asking for unnecessary identifiers. For sensitive uses, review the required contracts and safeguards before sending that information to any provider.
The Niwa setup guidance and safety information give product context. Check the configured request path with fictional test data and document what leaves your hosting. Do not infer a verified masking or deletion feature from an instruction telling the assistant to protect privacy.
Contact Niwa about the connection you intend to use and include the customer task, provider account type and data involved. Existing WordPress owners can use the standalone plugin; Niwa Business is USD 150 for a custom site or migration, hosting, maintenance and AI adaptation.
Provider terms checked on 9 October 2026. Reconfirm them for the selected account and feature before deployment.
