Atlassian started training AI on your Jira data this month. Most of you couldn't say no

There's a particular kind of policy update notice everyone deletes without reading, right up until the one time it quietly changes who gets to learn from your sprint board. Atlassian sent one of those in April. The deadline landed on August 17, ten days before this post went up, and a lot of Jira admins are only now finding out what they agreed to by doing nothing at all.

Starting August 17, 2026, Atlassian began collecting data from Jira, Confluence, and its other cloud products to train its AI features, including Rovo and Rovo Dev. That's a reversal of the company's prior stated position that customer content would not be used to train or improve its AI services. It now affects roughly 300,000 organizations, and how much say you get over it depends entirely on which plan you're paying for.

What actually changed

Atlassian splits the data into two buckets. The first is metadata: readability and complexity scores for Confluence pages, task classifications like "sales work item," semantic similarity scores between pages, story points on a Jira issue, a sprint's end date, an SLA value on a service request. The company says this is de-identified and aggregated before it touches any model, with names and email addresses stripped out.

The second bucket is in-app data, which is just the stuff your team actually wrote: Confluence page titles and bodies, Jira issue titles, descriptions, and comments, plus custom workflow and status names. For Free and Standard customers this is on by default, and it can be switched off. For Premium and Enterprise it starts off.

Metadata is the one that stings. Arseny Tseytlin, Atlassian's head of product communications, told The Register plainly: "If an Atlassian customer's highest active plan is Free, Standard, or Premium, metadata contribution is always on, and they are not able to opt out." Only Enterprise customers get a full switch for both categories, and Atlassian says it will hold everything it collects for up to seven years.

The part that should bother you even if you've never opened Jira

GitLab's response post to the policy put the underlying problem better than most: data protection is now a purchasing decision. Enterprise pricing at Atlassian starts around 801 seats with custom quotes attached, which is not a tier a five person agency or a mid sized product team backs into by accident. If your org sits below that line, the choice isn't "opt in or opt out." It's "pay a lot more, or don't get a choice."

That's a genuinely different posture than a cookie banner. A cookie banner is annoying and you click through it in two seconds. A metadata toggle that only exists once you've cleared a five or six figure spending threshold isn't consent, it's a loyalty program with privacy as the reward tier.

"De-identified" is doing a lot of work in that sentence

Atlassian's safeguards are real and worth stating plainly: names and emails stripped, data aggregated at a customer level, an explicit commitment to purge in-app data within 30 days of an opt-out and retrain affected models within 90. Very few companies bother writing a retraining timeline into a policy at all, and that's the kind of detail that separates a considered privacy program from a hand wave.

But GitLab's counterpoint holds too: story point patterns, sprint velocity, delivery cadence, and SLA breach rates are not sensitive in isolation, yet in aggregate they describe exactly how fast a team ships and how often it misses. "De-identified" tells you a name got scrubbed. It doesn't tell you whether the shape of the data underneath is still recognizable as yours once someone with the right dataset goes looking.

The honest counterpoint

It's worth crediting what Atlassian got right here, because the instinct to treat every AI training story as a scandal wears thin fast. There are real carve outs: customers on customer managed keys, Atlassian Government Cloud, Isolated Cloud, or with HIPAA requirements are excluded from collection entirely. And Atlassian isn't acting alone. GitLab's own post points out that GitHub recently shifted its Copilot data policy in a similar direction, which suggests opt-out-by-default is becoming the default posture across the industry rather than one company's overreach.

The stated goal, better search relevance inside Confluence, more accurate summaries, agentic workflows that actually complete a multi-step task instead of stalling halfway, is a legitimate one, and none of it happens without training data from somewhere. A company that wants those features and refuses to feed any model anything is asking for a car with no fuel.

The fair criticism isn't that Atlassian trains on data. It's that the default sits on "collect" for most of its customer base, and the honest opt-out lives behind a price tag most of those customers can't reach.

Why we build the other way

We've written before about what it actually means to own your data rather than just hold a copy of it, in Digital Sovereignty and in the recap of Local-First Conf 2026. Axtio's free tier keeps every card in your browser's own storage and never sends it anywhere, which means there's no metadata toggle to find because there's no collection happening to toggle off. Pro adds sync for people who want it, and the export is plain JSON either way.

We'll admit the comparison only goes so far. A personal action board isn't Jira running an engineering org's entire delivery pipeline, and nobody is building a foundation model off your grocery errands and the card you keep forgetting to move out of Other. The stakes here are smaller by design. That's rather the point of a tool built for one person instead of three hundred thousand companies: there's less of you to monetize twice.

If you actually run Jira or Confluence

The August 17 date has passed, but the settings are still live and still yours to change: Atlassian Administration, then Security, then Data contribution. Turning it off today still triggers the 30 day removal and 90 day retraining window Atlassian committed to. Nothing about missing the original deadline locks you out of using it. The only thing that's gone is the ability to have made this decision before the collection started rather than after.

Six weeks from now Microsoft is switching off Project Online and taking its customers' hosted data with it. Atlassian's story is the quieter version of the same lesson: the risk with hosted software was never only that it might disappear. Sometimes it just starts learning from you, and asks permission after the fact rather than before.

Sources

The Register, "Atlassian to train AI on user data unless law or cash say no", April 18, 2026. GitLab, "Atlassian will train on your data: Opt out with GitLab", May 4, 2026. Figures accurate as of August 2026.