Whose data trains what, and where does your captured material go?

What you are handing over is the part nobody can copy

Three hours of recorded interviews holds more of your business than any document you own. The cases, the numbers, the arguments you have refined over 20 years, the things you would not say on a stage.

So the question is fair, and it deserves an exact answer rather than a reassuring one: does that material end up inside a model that later serves someone else in your field?

The answer is no, on two conditions you can verify. The material has to sit under commercial terms rather than consumer terms, and it has to stay inside a workspace built for your brand alone. Both are checkable in writing, and most of the confusion around this question comes from mixing up two different things.

The workspace and the training run are not the same thing

A project is a workspace: a set of files, a set of instructions, and the chats that happen inside it. Your guides and transcripts live there and get retrieved to answer questions asked in that workspace. Anthropic's own documentation makes the boundary explicit at the level of individual chats, noting that context is not shared across chats unless it has been added to the project knowledge. A separate project is a separate set of files.

Training is a different process entirely. It happens at the level of the model, on the provider's schedule, using data the provider is permitted to use. Nothing about putting a file in a workspace puts it into a training run.

Once you hold those apart, the question stops being philosophical. It becomes a question about which contract governs your account, which is what a trained system is built inside.

What the two major providers say, with dates

Anthropic changed its consumer policy on August 28, 2025. Under the updated consumer terms, chats and coding sessions on the Free, Pro, and Max plans are used for model improvement unless the user opts out, and the retention window extends to five years for those who allow it, against 30 days for those who do not. Users had to make the choice to keep using the product.

The same announcement states that these updates do not apply to services under the commercial terms, which cover Claude for Work Team and Enterprise plans, the API, and Claude through Amazon Bedrock and Google Cloud's Vertex.

OpenAI splits the same way in the same direction. Its enterprise privacy page states that data from ChatGPT Business, Enterprise, Edu, and the API platform is not used for training by default unless the customer explicitly opts in. For individual accounts, the company's policy page is equally plain: content from consumer products may be used to train its models unless the user turns the setting off.

One brand name covers both arrangements. That is the whole trap. When someone tells you they use Claude or ChatGPT for your work, they have told you nothing about what happens to your material.

Deletion is a policy, and a court can suspend it

This is the uncomfortable part of the answer, and the part most vendors will not raise.

In May 2025, in the copyright litigation brought by news organizations, a magistrate judge ordered OpenAI to preserve consumer and API output logs, including conversations users had deleted. OpenAI opposed the order and lost. The company later confirmed that its obligation to retain that content ended on September 26, 2025, and its account of the dispute notes that ChatGPT Enterprise was excluded from the preservation order and that business customers with zero-data-retention agreements were unaffected.

The story did not end there. On January 5, 2026, a district judge affirmed orders requiring OpenAI to produce 20 million de-identified conversation logs to the plaintiffs, rejecting the argument that producing them would invade users' privacy. The people whose conversations sit in that sample were not asked.

The lesson is narrow and worth stating precisely. A retention policy describes what a company does on its own. Litigation can override it. The accounts that stayed outside the sweep were the commercial ones, which is the same line that governs training.

Could my material surface in someone else's answer?

There are two possible routes, and only one of them is a live mechanism.

The first route is retrieval, and it does not go anywhere. Files in your workspace are read to answer prompts in your workspace. A different client's project has different files.

The second route is training, and it only exists if your material entered a training run in the first place. Researchers have studied what models retain from training data for years. A January 2026 paper in Nature Communications found that models memorize more than exact repetition would predict, assembling fragments from similar sequences rather than only from identical ones, with repeated content carrying the highest risk.

Read that carefully, because it cuts both ways. Any single set of interview transcripts, sitting among trillions of words, is very unlikely to be reproducible from a model. Unlikely is doing work in that sentence, and it is why the sound approach is to keep the material out of training runs rather than to rely on the odds.

Four questions that settle it

Ask these of any marketer, agency, or consultant who will handle your recordings. They belong on the same list as everything else you check before signing.

• Which account holds my material, and is it governed by consumer terms or commercial terms? A provider who cannot name it has not thought about this.
• What written agreement says my content is not used to train models? Point me to the clause.
• Where do the recordings and transcripts live, how long are they kept, and who else can open them?
• If we stop working together, what happens to the recordings, the transcripts, and the workspace?

Ask before the first interview, because that is the session where the valuable material leaves your head.

My answers to those four questions

It would be poor form to hand you that list and then dodge it, so here is how the work runs here.

Your recordings sit in my Otter.ai account or my email, depending on how the session was captured. Your transcripts and guides sit in the Claude project built for your brand, which is the workspace your assets are produced from and holds nothing from anyone else's business.

Both are kept during the engagement and after it. That is deliberate rather than careless: a client who comes back two years later, or who wants a book built from material captured in year one, needs the original recordings to still exist. Material that has been deleted for tidiness cannot be used again.

There are no plan tiers here. Every engagement is built for the individual client, so there is no cheaper version of the work with looser handling attached to it. What is described in this section is how every engagement runs.

No client has asked me for a confidentiality agreement. If you want one, say so and I will draw one up and sign it before the first interview. That is a reasonable thing to ask for and an easy thing to give.

What follows you out the door

Your permanent property is not only the brand guides, although those matter most: they are documents in a format any future contractor or employee can read and work from on their first day.

It extends further. The transcripts are yours. The project itself is yours. I keep it accessible and I will export everything needed to run it, on request, whenever you ask.

In practice, most clients would rather I keep operating it, which is a preference rather than a lock. The distinction matters, and it is the one worth checking with anyone you hire. A vendor who will hand you the material is a vendor you are staying with by choice.

That is the whole answer. Your captured material serves one brand. It lives in a workspace built for you and nowhere else. The guides, the transcripts, and the project leave with you when you ask. And every sentence in this section is one you should make the person you hire say out loud before you record a word.