PMaker home
Search
中EN
System
Color theme
空格的键盘
空格的键盘空格的键盘产品经理 · AI Builder
The same sentence goes to different places depending on the tier you're usingWeb appConsumer products, loosest policy: conversations may improve the model, human reviewers may sample themMost convenient for personal use, worst place for sensitive materialAPIDeveloper access; most mainstream providers don't train on it by default, but retention and encryption details varyThe default choice for building AI features; terms differ by providerEnterpriseCommercial agreements: explicit no-training promise, NDAs, agreed data residencyThe only tier where data boundaries go into a contract; near-mandatory for complianceCheck the terms of the exact product you use—don't generalize

The same sentence goes to different places depending on which tier you're using. Figure out your tier before you talk data safety.

Where Conversation Data Goes

Web, API, and enterprise tiers route your data to different places.

Web, API, and enterprise tiers route your data to different places. "Where does my conversation data go?" is not a one-line answer—it depends on which form you use. Even within a single provider, the promises for the web app, the API, and the enterprise product can be completely different.

Signs you'll recognize:

  • A teammate pastes customer information into the web app and worries about it afterward.
  • The design doc says "data won't be used for training," but no one has checked the actual terms.
  • You need to ship a compliance-safe AI feature and aren't sure which integration tier to pick.

The three tiers

Web app (consumer). The free or subscription products aimed at regular users. Their data policies are usually the loosest: conversations may be used to improve the model, and human reviewers may see some of them. It's the most convenient tier for personal use and the worst place to paste sensitive information.

API (developer access). You call the interface to build your own application. This tier is typically not used for training by default—most mainstream providers work this way—but details like retention periods and encryption in transit still vary. OpenAI's enterprise privacy page, for example, states that API data after March 2023 is not used for training by default and that API inputs and outputs are retained for up to 30 days. openai-data This is the default choice for most "build an AI feature" scenarios. Which provider and channel you pick changes the terms.

Enterprise (commercial agreement). The tier for organizations: an explicit promise not to train on your data, the option of a non-disclosure agreement, and agreed data residency. It costs the most, and it's the only tier where you can put data-use boundaries into a contract. For customer data, healthcare, or finance, this is nearly mandatory. Anthropic's privacy policy likewise lays out consumer and commercial services separately, so you can check which one matches your integration. anthropic-privacy

A common mistake is assuming one company has one policy. Wrong. The same company can make very different commitments across its three tiers. Look up the terms for the exact product you're actually using; don't generalize.

How to read the terms

When you read the fine print, chase three questions.

One: will the data be used for training? This is the core clause. Watch whether the wording says "by default" or "promise." "By default not used for training" means you may need to opt in or out yourself; "promised not to be used for training" is a contract-level guarantee.

Two: how long is the data kept? How long are conversations stored, when are they deleted, and is there a delay window before deletion? As the product owner, you should know how long user data lingers behind your back.

Three: can humans see it? Some tiers note that human reviewers may sample conversations. That's a fatal detail for privacy-sensitive use, and it's often buried near the end of the terms.

Also, providers update their terms. You read them once at launch; six months later they change, and your compliance assumptions go stale. Put "review the terms periodically" into your maintenance routine.

As the product side

If you're building an AI product, you carry three responsibilities.

One: tell users plainly. Whether conversations get used for training, how long data is kept, and how to delete it—these belong in your privacy policy, written so a normal person can follow them. Moving users from guessing to knowing is the baseline.

Two: don't decide for your users. If you pick a channel that trains on data, you're handing user data to a third party for training—something the user never agreed to. Before you make that call, confirm your terms allow it and your users know about it.

Three: minimize data. Don't send what isn't needed. Send only the part required to complete the task and keep the rest in your hands. That's part of deciding where the boundary sits.

And the red line: once data enters third-party training, it's nearly impossible to withdraw. Files can be deleted; what's learned into weights cannot. Looking at the road before you send matters far more than repairing the damage afterward.

References

  1. Enterprise privacy at OpenAI
  2. Privacy Policy — Anthropic