19 of 25 paid AI business plans and APIs say in the contract they don't train on your data.
We read the published terms of 11 AI model developers, plan by plan, from ChatGPT and Claude to Gemini, Copilot and Grok. Every consumer plan trains by default.
- 11
- AI model developers
- 42
- plans, read one by one
- 104
- published documents, each read twice
Key findings
- On 19 of 25 paid business plans and APIs, the binding terms say the developer doesn't train on customers' data by default. 12 of those promises cover all of it. 7 cover only some models, such as Microsoft's foundation models for Azure OpenAI and Microsoft 365 Copilot.
- 5 paid business plans train by default: Mistral's API and Team plan, Cohere's API and North, and DeepSeek's API.
- Amazon Bedrock makes the promise only in its FAQ, which isn't part of the contract.
- All 12 consumer plans we read train on users' chats by default, from ChatGPT and Claude to Gemini, Copilot and Grok, and all of them let users opt out.
- The 5 free or discounted developer plans we read train by default too.
- Since mid-2023, 3 of the 5 consumer plans we could compare moved toward training, and no business plan did. In 2023, Anthropic's terms promised not to train on Claude.ai chats. Its consumer plans now train by default, with an opt-out.
Business plans and APIs.
On 19 of 25, not training on customers' data by default is in the binding terms. Each plan links to the clause.
- Amazon BedrockAmazon, business plan or APIDoesn't train by defaultOn a web page only
- Claude APIAnthropic, business plan or APIDoesn't train by defaultIn the contract
- Claude EnterpriseAnthropic, business plan or APIDoesn't train by defaultIn the contract
- Claude TeamAnthropic, business plan or APIDoesn't train by defaultIn the contract
- Cohere APICohere, business plan or APITrains by default, with an opt-outNo promise
- NorthCohere, business plan or APITrains by default, no opt-outNo promise
- DeepSeek APIDeepSeek, business plan or APITrains by default, with an opt-outNo promise
- Gemini API, paidGoogle, business plan or APIDoesn't train by defaultIn the contract
- Gemini EnterpriseGoogle, business plan or APIDoesn't train by defaultIn the contract
- Gemini in Google WorkspaceGoogle, business plan or APIDoesn't train by defaultIn the contract, Workspace AI models only
- Vertex AIGoogle, business plan or APIDoesn't train by defaultIn the contract
- Meta Model API, standardMeta, business plan or APIDoesn't train by defaultIn the contract, Meta's models only
- Azure OpenAIMicrosoft, business plan or APIDoesn't train by defaultIn the contract, foundation models only
- Microsoft 365 CopilotMicrosoft, business plan or APIDoesn't train by defaultIn the contract, foundation models only
- Mistral APIMistral, business plan or APITrains by default, with an opt-outNo promise
- Vibe EnterpriseMistral, business plan or APIDoesn't train by defaultIn the contract, preview models left out
- Vibe TeamMistral, business plan or APITrains by default, with an opt-outNo promise
- ChatGPT BusinessOpenAI, business plan or APIDoesn't train by defaultIn the contract
- ChatGPT EnterpriseOpenAI, business plan or APIDoesn't train by defaultIn the contract
- OpenAI APIOpenAI, business plan or APIDoesn't train by defaultIn the contract
- Enterprise Pro and MaxPerplexity, business plan or APIDoesn't train by defaultIn the contract, generative AI models only
- Sonar APIPerplexity, business plan or APIDoesn't train by defaultIn the contract, generative AI models only
- Grok BusinessxAI, business plan or APIDoesn't train by defaultIn the contract
- Grok EnterprisexAI, business plan or APIDoesn't train by defaultIn the contract
- xAI APIxAI, business plan or APIDoesn't train by defaultIn the contract
fulkersonadvisors.com
Consumer and free plans.
All 12 consumer plans and all 5 free or discounted developer plans train by default. Each plan links to the clause.
- Claude FreeAnthropic, consumer planTrains by default, with an opt-outNo promise
- Claude Pro and MaxAnthropic, consumer planTrains by default, with an opt-outNo promise
- Cohere API, trialCohere, free or discounted APITrains by default, no opt-outNo promise
- DeepSeek appDeepSeek, consumer planTrains by default, with an opt-outNo promise
- Gemini API, free tierGoogle, free or discounted APITrains by default, no opt-outNo promise
- Gemini app, personal accountGoogle, consumer planTrains by default, with an opt-outNo promise
- Meta Model API, contributorMeta, free or discounted APITrains by default, no opt-outNo promise
- Muse CodeMeta, free or discounted APITrains by default, with an opt-outNo promise
- Copilot, consumerMicrosoft, consumer planTrains by default, with an opt-outNo promise
- Mistral API, free modeMistral, free or discounted APITrains by default, with an opt-outNo promise
- Vibe FreeMistral, consumer planTrains by default, with an opt-outNo promise
- Vibe ProMistral, consumer planTrains by default, with an opt-outNo promise
- ChatGPT FreeOpenAI, consumer planTrains by default, with an opt-outNo promise
- ChatGPT Plus and ProOpenAI, consumer planTrains by default, with an opt-outNo promise
- Perplexity FreePerplexity, consumer planTrains by default, with an opt-outNo promise
- Perplexity Pro and MaxPerplexity, consumer planTrains by default, with an opt-outNo promise
- Grok, consumerxAI, consumer planTrains by default, with an opt-outNo promise
fulkersonadvisors.com
Since mid-2023, 3 of 5 consumer plans here moved toward training. No business plan did.
What the same developers' documents said before ChatGPT Enterprise launched, from the Internet Archive's copies, against what they say now. In 2023, Anthropic's terms promised not to train on Claude.ai chats. Each name links to its document.
Point at or tap a line to read what the documents said.
fulkersonadvisors.com
Legal AI vendors mostly don't.
Only 4 of 29 legal AI vendors put a full promise not to train on client data in their own contract. Vendor by vendor.
Know what your AI vendors can do with your data.
We build AI systems for companies and read the vendor terms that come with them. Bring yours to a thirty-minute call.
How we did it
We read the published documents that govern customers of 11 AI model developers, plan by plan, 42 plans in all: terms of service, business and API terms, data processing agreements, privacy notices, trust and help pages, 104 documents in all, on October 12, 2026.
An AI model (Jev, from TypeSafe) marked every passage about training, opting out, feedback, retention and changes to the terms, then answered each question for each vendor and plan, pointing to the passage behind each answer. A second AI reader read the same documents in full without seeing those answers. The two agreed on the default and on where the promise sits for 32 of 42 plans. Every difference was settled by reading the clause, with a written reason, and every quote in the data was checked word for word against its document.
A promise counts as in the contract only when a binding document (terms, an enterprise or AI agreement or a data processing agreement) says the vendor won't train on customers' data. A promise that covers only the outside AI companies a vendor uses doesn't count as the vendor's own. Where documents disagree, the binding one decides.
For OpenAI, Anthropic, Google, Microsoft, Cohere and Perplexity we also read what their documents said from January 1 to August 27, 2023, before ChatGPT Enterprise launched. We listed every page the Internet Archive captured in that window under each developer's legal, privacy, help and docs pages (9,904 pages), kept the 504 that could hold terms or data policies, and read the 525 readable copies, the first and last of each in the window. The same AI model and a second reader answered for each product, and every difference was settled by reading the clause. Anthropic's 2023 privacy policy, which isn't binding, also listed de-identifying prompts and outputs to train its models. Its terms decide, as for every vendor. No Anthropic copy before June 2023 survives. Amazon Bedrock was in preview with no published document on the question, so it isn't compared.
These are published terms. Large customers often sign their own agreements. Three plans were left out: Amazon Q Business, which no longer takes new customers, and Amazon Q Developer and Microsoft Copilot Studio, whose governing terms we didn't capture. The same reading of 29 legal AI vendors is a separate study.
Cite this study
Fulkerson Advisors (2026). Do AI companies train on your data? The published terms of 11 AI model developers, plan by plan. https://www.fulkersonadvisors.com/research/ai-companies-training-on-business-data
Free to use with a link to this page. The data is in the CSV.
Questions
Does ChatGPT train on my data?
On ChatGPT Business, ChatGPT Enterprise and the API, OpenAI's terms say it doesn't train on customer content by default. ChatGPT Free, Plus and Pro train on conversations by default, with an opt-out.
Does Claude train on my data?
On Claude Team, Claude Enterprise and the API, Anthropic's commercial terms say it doesn't train on customer content by default. Claude Free, Pro and Max train on chats by default, with an opt-out. In 2023 its terms promised not to train on Claude.ai chats.
Does Gemini train on my data?
On Vertex AI, the paid Gemini API, Gemini Enterprise and Gemini in Google Workspace, Google's terms say it doesn't train on customer data by default. The Gemini app on a personal account trains by default, with an opt-out, and the free tier of the Gemini API trains with no opt-out.
Does Microsoft Copilot train on my data?
Microsoft's product terms say Azure OpenAI and Microsoft 365 Copilot don't use customer data to train the foundation models. The consumer Copilot app trains by default, with an opt-out.
Which AI companies train on paid business data by default?
Of 25 paid business plans and APIs we read, 5 train by default: Mistral's API and Team plan, Cohere's API and North, and DeepSeek's API.
Is "no training" the default on business AI plans, or negotiated?
With the big model developers it is the default, and it is in the contract on 19 of 25 paid business plans and APIs. Every consumer plan we read trains by default.