Atlassian now trains Rovo AI on Cloud customer content by default, with full opt-out Enterprise-only

What changed on 17 August

According to a report on dev.to, originally published at theaidownside.com, Atlassian began feeding content from its Cloud products into the training pipeline for Rovo, its AI assistant, on 17 August. The change was opt-out rather than opt-in: Confluence pages, Jira work items and the descriptions and comments attached to them became training material by default, leaving customers to locate a setting if they wanted to stop it.

Atlassian's own data-contribution documentation, as cited by the report, confirms the arrangement, and the defaults differ by plan tier in ways that matter.

Two settings, unequal control

There are two separate data-contribution settings, and per the dev.to report they are not equivalent. One covers in-app data — the text users actually write. The other covers metadata — derived signals about that text.

On the Free, Standard and Premium plans, an administrator can switch off in-app data contribution but cannot disable metadata contribution; the report quotes Atlassian's support page stating plainly, "You can't change this setting." Only Enterprise customers receive the switch that also stops metadata sharing.

The defaults vary too. In-app data contribution is enabled by default for Free and Standard customers and disabled for Premium and Enterprise, though every tier can toggle it. Metadata contribution is on across the board and can only be turned off on Enterprise. The net effect, the source argues, is that the customer on the cheapest plan contributes the most by default and has the fewest means to stop it.

What data is covered

Per Atlassian's materials as described by dev.to, in-app data includes Confluence page titles and body text, Jira work-item titles, descriptions and comments, and custom status and workflow names. Metadata covers the derived layer: readability scores, task classifications (such as labelling a ticket as sales work), story points, sprint end dates, SLA values and semantic-similarity measures drawn from Atlassian's Teamwork Graph.

None of that metadata is raw prose, but as the report notes, a great deal can be inferred from it — how a team works, what it is working on, how fast it moves and how everything connects.

Retention and the company's stated position

The report lays out Atlassian's side. Rovo competes in a market where an assistant tuned to a customer's workspace beats a generic one, and that understanding is built from customer data. Premium and Enterprise tiers received a gentler default for in-app content. Retention is bounded on paper: after an opt-out, in-app data leaves improvement datasets within roughly 30 days and content attributes within roughly 90, with retraining to follow. Aggregated, de-identified data may persist for up to seven years.

The source also raises a harder question underneath that timeline: removing a document from a training set does not erase the influence it had on a model that already learned from it. Given the unresolved difficulty of machine unlearning, a promise to retrain is a commitment to hold the vendor to rather than accept on trust.

What administrators can do

The opt-out path is documented and not especially buried. An organisation administrator goes to Atlassian Administration, then Security, then Data contribution, and disables in-app data contribution; on Enterprise, metadata contribution can be switched off as well.

The report's structural criticism is that this design shifts the burden onto customers. An admin has to know the change happened, know the setting exists, and act before the default does the contributing. Individual users writing tickets and pages have no switch at all — their data's fate is decided a level above them. And on plans where metadata cannot be disabled, even an admin who flips every available toggle is still contributing something.

Why it matters

The story fits a broader pattern of platforms treating user content as training data by default, with Twitch's automatic enrolment of streamers into Amazon's AI training cited as a parallel. The workplace-tools version is arguably stickier, because a decade of Jira history is not something a team can migrate as casually as switching chatbots.

The specific objection raised by the source is narrow but pointed: not that a software company built an AI feature, but that a consequential privacy change was made the default, that the completeness of control over it is metered by price, and that the customers with the least protection are the ones paying the least. If the tiering stands, privacy becomes a line item on the contract — and as the rest of the industry drifts the same way, that is the design decision worth watching.