The question people ask is “does it train on my data”. The question that decides anything is narrower: which plan, which product, and is the answer a setting or a contract.
The line falls at the plan, not the vendor
No major vendor treats all its customers the same way. ChatGPT uses consumer conversations for training by default with an opt-out in settings, and excludes Team and Enterprise workspaces contractually. Claude does the same: consumer plans train unless you opt out, Team, Enterprise and API traffic are excluded. Gemini goes further on the consumer side, where a sample of conversations is reviewed by humans and kept up to three years unless you turn off Gemini Apps Activity, while Workspace accounts are excluded.
So “does Anthropic train on my data” has no answer. “Does Claude Team train on my data” has one, and it is no.
A setting and a contract are not the same protection
An opt-out toggle can be reset by a product update, is set per account rather than per organisation, and depends on every person on your team having found it. A contractual exclusion on a business plan applies to the workspace whether or not anyone touches a setting.
For a company, that difference is the reason to move people off expensed individual seats and onto a team plan, more than the shared billing.
Some products never train on your content at all
A second group excludes customer content by design rather than by tier.
Microsoft 365 Copilot states tenant content is excluded from foundation model training, with prompts and responses kept under existing Microsoft 365 retention policies and an option to stay inside a European data boundary.
DeepL deletes submitted text immediately after translation on paid plans and never uses it for training. Its free translator does not carry that promise, which is the single most important line on the page for anyone translating a contract.
Adobe Firefly approaches it from the other end: its own models are trained on Adobe Stock and licensed content rather than on customer work, and enterprise customers get IP indemnification. Note the boundary, though, since output from the third-party partner models available inside Firefly follows those vendors’ terms instead.
Training is not the only question
Three things get conflated on vendor pages, and they are independent.
Training is whether your content improves a model other people use. Retention is how long it sits on the vendor’s systems. Human review is whether staff or contractors read samples of it. A product can exclude your content from training and still store it for years, and still have someone read it for quality.
Gemini’s consumer terms are the clearest example: activity off stops the training use, and the retention and review questions have their own answers.
How to read a plan in two minutes
Look for four sentences, in this order.
Is customer content used to train models, and does the answer depend on the plan. How long is content retained, and can it be deleted on request. Is there human review, and by whom. Where is the data stored, if a region matters to you.
If a vendor answers only the first, that is itself an answer.
Every listing in this directory carries those terms in its privacy section, taken from the vendor’s own policy and dated. The ones where the answer is a clean no are the enterprise plans, where it is written into the contract rather than into a settings page.
