Your AI Vendor's Data Policy Keeps Changing. It's the Same Gap Five Times.

If you've ever flipped a privacy toggle, skimmed a vendor's retention policy, or decided an AI tool was "fine" because the marketing page said so, you made a one-time decision about something that keeps changing without telling you. Five different vendors proved that this year, five different ways. The check that survives all five isn't a setting — it's three questions you ask on a schedule, not once at sign-up.

Five vendors, one blind spot

Meta's updated business terms let it scrape content tied to any business account for AI training once the October 30 deadline passes, with no opt-out. ChatGPT's "Improve the model for everyone" toggle was reported reverting to on after app updates — quietly, no dialog, no email. A personal AI assistant called Instinct kept full copies of users' emails after they revoked account access; disconnecting stopped new reads, not the mail already copied. A free, on-device PII filter OpenAI released catches 96% of personal identifiers before they reach a server — and the 4% it misses is still the team's problem, not the tool's. And Atlassian's updated terms put every Free, Standard, and Premium customer's Jira tickets and Confluence pages into AI training by default, with full opt-out locked behind Enterprise pricing.

The pattern underneath

None of the five teams in those stories did anything careless. They read the terms that existed when they signed up, flipped the toggle that existed then, and trusted that "opted out" and "disconnected" meant what they sound like. What changed wasn't their judgment — it was the ground under the setting. A privacy policy, a retention window, and a training default aren't fixed facts about a tool. They're product decisions a vendor can revise, and revise again, usually without a change log a regular user will ever see.

That's the mechanism under all five stories: the setting you checked and the setting actually live right now are two different things, and only the vendor knows for sure which one is in effect this week.

What to do differently Monday morning

Stop treating any AI tool's privacy setting as something you confirm once at sign-up. For every tool that touches client data, financial figures, or anything you'd rather a competitor not read, get three specific answers — and recheck them on a schedule, not just when a headline makes you nervous:

1. Does it train on your input, and can you actually opt out? Not whether there's a toggle — whether the toggle survives an app update, and whether opting out exists at your plan tier or sits behind a higher one.
2. What happens to data you've already shared if you disconnect or opt out? A setting that stops future collection isn't the same as deleting what's already been copied.
3. Where does the remaining risk sit? A PII filter, a redaction step, and a retention promise each cover part of the problem. None covers all of it, and knowing which gap is still yours to watch is the actual skill.

Priya: auditing a tool before she trusts it with anything new

Priya runs operations at a 60-person logistics company evaluating a new AI scheduling assistant for the team's shared calendar and email. The vendor's page says "enterprise-grade security" and nothing more specific. Before connecting anything, she asks support three direct questions: how long is calendar data retained, where does it live, and if she disconnects in six months, does it delete what the tool already read or just stop reading new events. The answer to the third question is a vague "we'll look into that." That's the answer she uses to decide, not the marketing copy.

Dana: catching a setting that moved without her

Dana manages client communications at a 25-person PR firm and opted her team's ChatGPT accounts out of model training back in the spring, specifically because client campaign drafts reference unreleased product launches. She hadn't checked the setting since. Running the same three questions again this month, she found two of four accounts had reverted to the default after an app update — the reset reported elsewhere this year — and nobody had noticed, because nothing in the product told them it happened. She didn't need a new tool. She needed the recheck on a calendar, not a one-time setup step.

The one thing

Five vendors made five different calls on your data this year, and the common thread isn't which one is most trustworthy. It's that all five calls can change again without telling you. The setting you checked at onboarding is a snapshot, not a guarantee. The habit that holds up is checking it again, on your own schedule, before the next vendor makes its own call for you.