Think about the last time you rented an apartment and found a clause on page nine saying the landlord could walk through your living room whenever they felt like redecorating. You’d notice. You’d push back. You’d at least know the clause existed.
AI training policies work more like page nine, except the walkthrough already started and the notice arrived as a changelog entry. I spend my days testing agents and the tools they plug into, and the pattern I keep hitting is not that companies refuse to let you opt out. Most of them do. The problem is that the opt-out lives somewhere you’d never think to look, and the clock is already running.
Three Companies, Three Very Different Answers
Take the current state of play across a few platforms I actually use.
Mistral has the cleanest version. Once your opt-out is confirmed, Mistral no longer uses your input or output data to train its models. Input and output, both directions, stated plainly. That is the standard everything else should be measured against.
GitLab also offers an opt-out. Fine. Straightforward.
Then there’s GitHub Copilot, where the situation gets murkier. Copilot may use private repository data without a prior opt-out, and the language in the privacy update covers interaction data specifically inputs, outputs, code snippets, and associated context that will be used to train and improve their models. Meaning if you used Copilot inside a private repo, that context traveled with it. Developers digging through their settings to turn this off have reported that the Privacy section only surfaced “Suggestions matching public code” and no model training toggle they could find. When people who write software professionally cannot locate the switch, the switch is not really being offered.
And Atlassian has put a date on the calendar. Starting August 17, 2026, Atlassian begins collecting customer metadata and in-app content from Jira, Confluence, and other cloud products to train its AI offerings, including Rovo and Rovo Dev. If your team runs its planning, retros, incident writeups, and half-finished architecture arguments in Confluence, that is the material in question.
Why Default-On Is the Whole Story
Every one of these policies could be identical in legal substance and still land completely differently based on one design choice: whether the default is on or off.
Opt-out means the company has your data unless you act. Opt-in means they don’t have it unless you act. That single flip determines what happens to the overwhelming majority of users, because the overwhelming majority of users never open the settings page. Product teams know this. It is not an accident that the harder-to-find toggle is the one that costs the company training data.
Regulators are starting to poke at this. Under the EU AI Act, an opt-out trend is emerging that may limit how data gets used for training, though it is not yet clear how a legally valid opt-out request would actually be made. Collecting societies including the Dutch organization Pictoright, which represents pictures, and the French society Sacem have been pushing on the question. The mechanism is unsettled. The direction is not.
What I Actually Do About It
My working approach, which I’d suggest to anyone running agents on real work:
- Check platform settings for opt-out options on every tool that touches your code, docs, or customer data. Not once. Every time a privacy policy update email arrives, because that email is often the only notice you get.
- Treat announced dates as deadlines, not suggestions. August 17, 2026 is a real date for Atlassian customers. Put it somewhere you’ll see it.
- Assume the setting is not where you expect. If you look in the obvious privacy panel and find nothing, that does not mean the toggle doesn’t exist. Search the docs, check org-level admin settings, and ask support directly.
- Distinguish input from output. Some policies cover only what you type. Mistral’s covers both directions. Know which one you signed.
- Make it an org decision, not an individual one. Individual developers toggling their own accounts does nothing for a shared Jira instance or a team’s private repos. Someone with admin access has to own this.
The Uncomfortable Part
None of this is a scandal. Companies training on usage data to improve products is old news, and the tools do get better for it. What bothers me is the asymmetry in effort. Turning the feature on takes them one policy update. Turning it off takes you an afternoon of clicking through settings menus that may not contain what you’re looking for.
The vendors that publish a clear statement, name a specific date, and put the switch somewhere findable are doing something genuinely useful. The ones that bury it are making a bet about how many of us will bother. Prove them wrong on your own account, then go do it for your team.
🕒 Published: