Privacy, Data, and Where Your Prompts Go

The concept

Every message you type into an AI tool goes somewhere, and understanding exactly where is the first real privacy skill in this course — not because AI companies are uniquely untrustworthy, but because the defaults vary by tool and by plan, and most people never check. Most consumer-facing chat tools may log conversations and, depending on account settings, use them to improve or train future models — unless you explicitly opt out in settings, or you're on an enterprise or API tier where the vendor's terms typically promise no training on your inputs as part of the contract. These are meaningfully different data-handling regimes sitting behind what looks like the same chat window, and the only way to know which one you're in is to actually check the settings, not assume based on how the product feels.

What counts as sensitive input is broader than people initially assume. The obvious categories — customer personal data, credentials, API keys — are easy to flag. Less obvious ones catch people more often: unreleased product plans or pricing pasted in for a "can you help me phrase this" request, an internal document with a client's name left in context, a draft contract with real counterparty details, or a screenshot containing a name and account number you didn't think to blur. Each of these, once pasted into a tool without a no-training guarantee, may persist in logs you don't control and can't delete on demand.

The practical rule that covers nearly every case: never paste anything into an AI tool you wouldn't be comfortable seeing in a vendor's logs. This isn't a reason to avoid AI tools for real work — it's a filter applied before you paste, not a reason not to paste at all. In practice this usually means one of three things: use placeholders for the sensitive specifics ("Customer A," "$X," a fake but structurally similar account number) and fill in the real values yourself afterward; use your organization's enterprise or API-tier access, if one exists, for anything genuinely sensitive; or, if neither is available, don't paste that particular piece of information at all and do that part of the task manually.

Checking your actual tool's settings is not a one-time task. Vendors change defaults, add new toggles, and ship new tiers; a no-training opt-out you set a year ago may not carry over to a new feature or a new device. Treat a periodic settings check — once a quarter is a reasonable cadence — as part of using AI tools responsibly, the same way you'd periodically review who has access to a shared drive.

Worked example

A learner compares two real settings screens side by side. The first is a consumer chat app's data controls page, where a toggle labeled "Improve the model for everyone" is switched on by default — turning it off stops that account's new conversations from being used for training going forward, though it typically doesn't retroactively remove what's already been logged. The second is an enterprise plan's terms-of-service excerpt, which states inputs and outputs are not used for training and are retained only for a limited period for abuse monitoring, as part of the paid contract rather than a toggle the user has to find and set themselves.

The learner realizes they had been using the free consumer version of a tool for draft client emails at work — with the training toggle left in its default "on" position — without ever having opened the settings page to check. They switch the toggle off for future use and flag to themselves that anything already pasted under the old default can't be un-sent, which is the real cost of not checking sooner.

Exercise

Open the actual data-handling or privacy settings of the AI tool you use most, and find the specific toggle or clause that governs whether your inputs are used for training and how long they're retained. Note what the default was when you checked (on or off) — most people have never looked.

Then, from memory or your recent chat history, review your last five real prompts to that tool. Flag any that included information you wouldn't want sitting in a vendor's logs indefinitely — a client name, an unreleased number, an account detail. Rewrite one of the flagged prompts using placeholders for the sensitive specifics, keeping the structure and request identical, and compare how much (or how little) the useful part of the prompt actually depended on the real data being present.

Recap

• Data-handling defaults vary by tool and by plan — check your actual settings rather than assuming; a consumer free tier and an enterprise tier can be fundamentally different regimes behind an identical-looking chat window.
• Sensitive input is broader than credentials and customer records — unreleased plans, pricing, and incidental details in pasted documents all count.
• The working rule: never paste anything you wouldn't be comfortable seeing in a vendor's logs — use placeholders, use a no-training tier, or do that piece manually instead.

Resources

• OpenAI: "Privacy policy" — https://openai.com/policies/privacy-policy
• Anthropic: "Privacy Center" — https://privacy.anthropic.com/
• NIST: "AI Risk Management Framework" — https://www.nist.gov/itl/ai-risk-management-framework