Blog
The privacy policy is probably not the document that applies to you
Sam Carr
If you are thinking about putting a tool in front of your customers' data, someone will eventually ask you the question every business owner dreads: does it train on our stuff?
The answer is usually written down. The problem is that most AI companies have written it down more than once, in documents that do not agree, and the one that turns up first in a search is generally the one that does not apply to you.
There are two documents, and only one of them is yours
Every large AI provider publishes a consumer privacy policy covering the chat product anyone can sign up to, and a commercial agreement covering the API and the business accounts. They are different products with different promises, and the commercial one is the document a tool built for your business runs under.
Anthropic's privacy policy says so itself:
This Privacy Policy does not apply to content that we process on behalf of customers of our business offerings, such as our Enterprise accounts.
That single sentence is the whole trap. Read the privacy policy on its own and you will find that Anthropic lists, among the sources it trains on, "Data that our users or crowd workers provide, including Inputs and Outputs from our Services (unless users opt out)". Which is true, and is about the consumer product, and is not the document your supplier would be operating under.
What the commercial contracts actually say
Checked on 26 August 2026, against each company's own published terms rather than a summary of them.
| Provider | Document | Clause | What it says |
|---|---|---|---|
| Anthropic | Commercial Terms of Service | B | "Anthropic may not train models on Customer Content from Services." |
| OpenAI | Services Agreement | 4.2 | "OpenAI will not use Customer Content to develop or improve the Services, unless Customer explicitly agrees to such use." |
| Perplexity | API Terms of Service | 2.3.3 | "Perplexity shall not use (or authorize third parties to use) Customer Content to train, retrain, fine-tune or otherwise improve any generative artificial intelligence models." |
Three things are worth noticing about that table.
They are contract terms, not settings. None of these is a checkbox somebody can leave switched off by accident. They bind the provider by default, for everybody on those terms.
The wording differs and the difference matters. Anthropic's is flat. OpenAI's carries an "unless Customer explicitly agrees" exception, so it is worth knowing what your supplier has agreed to on your behalf. Perplexity's is the broadest of the three, because it names retraining and fine-tuning as well as training, and it extends the promise to third parties.
Dates move. Each of those documents carries an effective date and each of them can be reissued. A quote with no date against it is not evidence of anything a year later, which is why the date above is written into the sentence rather than left to the reader to guess.
A promise in the documentation is not a promise in the contract
This is the distinction that does the most work and gets noticed the least.
A company can tell you something in three quite different places, and they are not equally binding:
- In the contract you are a party to. It is enforceable and it cannot be changed without notice.
- In a policy page. Usually versioned, usually dated, changeable by them.
- In the developer documentation. A help page. It can be edited this afternoon, and often has no version history at all to tell you that it was.
All three read identically when you are skimming. Only one of them is a promise. If the answer to "do they keep our data" lives on a documentation page rather than in the terms, what you have is a description of current behaviour, not a commitment to it.
Retention is where this bites most often, because retention is frequently answered in a different document from training. Finding a strong no-training clause tells you nothing at all about how long anything is stored, and the two questions have to be asked separately.
What to ask before you sign anything
Whoever is building the tool should be able to answer all of these without going away to look:
- Which provider is it, by name? Not "an AI model". The company whose terms you will be relying on.
- Which document covers us, and can I have the link? If the answer is the consumer privacy policy, that is the wrong document.
- What does the training clause say, word for word? A summary is not a clause.
- Where is retention answered, and is it in the contract or in the documentation?
- What happens to the data if you stop working with them?
None of these need a lawyer to ask. They need somebody who has actually opened the documents, and the answers should end up written down beside the scope and the price rather than in an email thread. How a build gets scoped and agreed is the right place for them, and it is worth asking to see that in writing before anything starts.
The short version
The AI companies have, on the whole, written down reasonable commitments for their business customers. What they have also done is publish a consumer policy that says something different, and leave the two a search away from each other.
So the useful habit is not to distrust the answer. It is to check which document the answer came out of, and to write the date on it. If you want to talk through a specific provider, get in touch.