Does Your AI Vendor Train on Your Data? A Buyer's Checklist for Retention and Opt-Ins
Anthropic and OpenAI both say they do not train their models on data sent through their business and API products by default. That answers the first question a buyer asks, but not the next three: how long the data is kept, which exceptions let it into training, and which features sit outside the strongest retention controls. This checklist is based on both vendors' own pages as read on 10 November 2026. Terms change, so confirm each point against your contract before you rely on it.
Key facts
- Anthropic's privacy page, dated 18 August 2026, says it does not use inputs or outputs from its commercial products to train models by default.
- The same page lists exceptions: content you report as feedback or bugs, or content you otherwise choose to allow. Thumbs up or down feedback is kept with the whole related conversation for up to 5 years, de-linked from user and customer IDs.
- OpenAI's data controls page says data sent to the OpenAI API is not used to train or improve its models unless you opt in.
- OpenAI generates abuse monitoring logs by default and keeps them for up to 30 days unless the law or protection from harm requires longer. Zero data retention needs prior approval from OpenAI.
- OpenAI's Responses API keeps application state for 30 days by default. Assistants and Threads objects not deleted through the API or dashboard are kept indefinitely.
Check which plan you are on
The statements above apply to commercial products. Anthropic's page says it does not cover its consumer plans (Free, Pro and Max) and refers readers to a separate article. The most common mistake in a small company is an employee using a personal account for work. A vendor's business terms protect only the accounts that sit under them. Ask which accounts your team actually uses, and move work to a managed plan.
Read the exceptions, not the headline
A no-training default is only as strong as its exceptions. Anthropic's exceptions are feedback you submit and anything you explicitly allow. Its page says admins on Team and Enterprise plans can turn off feedback submission, through the Rate chats setting under Data and Privacy. OpenAI's statement carries an opt-in condition. In both cases, the practical question is who in your organization can click a feedback button or opt in, and whether you have switched that off.
Ask about retention, which is a separate question
Not training on data does not mean not storing it. OpenAI documents several clocks. Abuse monitoring logs are kept up to 30 days by default. The Responses API keeps application state for 30 days by default, and setting store to true also keeps data for at least 30 days. Deleted Assistants and Threads objects are removed 30 days after deletion, while objects you never delete stay indefinitely. Anthropic's page, for its part, addresses feedback retention of up to 5 years and says other retention periods sit in separate articles we did not review for this guide.
When you compare vendors, ask for a retention table with one row per feature you plan to use, not one figure for the whole company.
Zero data retention is narrower than it sounds
OpenAI says zero data retention is subject to its approval and extra requirements. It also publishes a list of endpoints that are eligible, some with limitations, and a list that are not. Not eligible, per that page: Conversations, Assistants, Threads, Vector Stores, Files, Fine-tuning, Evals, Batches and Agents. If your design uses file search or stored threads, zero data retention would not cover those parts. The page also says image and file inputs are scanned for child sexual abuse material, and a flagged image is retained for manual review even with zero data retention enabled.
Data residency is not the same as retention
OpenAI states that customer content is stored at rest in the region you select for supported endpoints, and that regional inference is available in the US, the EU (EEA and Switzerland) and the UAE. It adds that data residency does not apply to system data such as account, usage and billing data. A European buyer should ask where account, usage and billing data live as well as where content lives.
What to do
- List every AI account your team uses and confirm each sits under a business or API agreement.
- Turn off feedback submission and any opt-in sharing where your admin console allows it.
- Request a per-feature retention table in writing, and keep it with your vendor file.
- If you need zero data retention, check each endpoint you use against the vendor's eligibility list before you design around it.
- Re-read the vendor's page on each contract renewal. Dates on these pages move.
For vendor assurance reports, see our note on SOC 2 audits of AI vendors, and for the cost side of governance gaps, what IBM's breach report says about AI governance. Browse listed vendors in the AI tool directory.
Questions people ask
Does Anthropic train on API data?
By default, no. Its privacy page says it does not use inputs or outputs from commercial products to train models, with exceptions for feedback you report and anything you choose to allow.
Does OpenAI train on API data?
Its data controls page says data sent to the API is not used to train or improve its models unless you explicitly opt in to share it.
What does zero data retention mean?
It is a vendor-approved setting under which eligible endpoints do not store your content after the request. On OpenAI it needs approval, and several endpoints, including Batches, Files and Vector Stores, are not eligible.
Sources