Atlassian Announces AI Training Data Collection—Free and Paid Users Can't Opt Out

doc

Atlassian announced on April 18-19, 2026 that starting August 17, 2026, it will collect customer metadata and in-app content from Jira, Confluence and other cloud products by default to train AI models (Rovo, Rovo Dev). Affecting roughly 300,000 global customers, the policy creates a controversial two-tier privacy system where only Enterprise users can opt out.

Starting August 17, 2026, Atlassian will use customer metadata and in-app data to improve the AI experience across Jira, Confluence, Jira Service Management, and platform applications such as Rovo. The policy splits the two data types by plan tier: Enterprise customers can opt out of metadata contribution, while Free, Standard, and Premium plans have metadata contribution enabled by default with no opt-out, and every customer can opt out of in-app data contribution. This piece breaks down what each category means, how removal works, and what procurement and security teams should ask before accepting the terms.

What the policy covers and when it takes effect

The change applies to the company's cloud products where AI features are offered, including issue tracking, documentation, service management, and the Rovo assistant. The stated purpose is improving the AI experience itself: responses, recommendations, and product behavior tuned on patterns drawn from how customers use the platform.

Two distinct inputs feed that improvement. Keeping them separate matters, because the opt-out rights and removal timelines differ between them, and because one category carries more identifying structure than the other.

Metadata versus in-app data

  • Metadata describes the structure and activity around the work: project types, workflow arrangements, usage patterns, configuration signals, and how people interact with the product. It often omits the actual written content, but it still carries information about how an organization operates.
  • In-app data is the content itself: issues, comments, pages, attachments, service requests, and assistant conversations held inside the applications.

Before contribution, the company applies de-identification and aggregation, removing names and email addresses from what enters the datasets. The distinction between the two categories is useful when reading the opt-out rules, because metadata can remain identifying in aggregate even after direct names are stripped.

Who can opt out, and who cannot

Plan tierMetadata contributionIn-app data contributionOpt-out position
EnterpriseContributed to improve AIContributed to improve AICan opt out of metadata; can also opt out of in-app data.
FreeAlways enabledContributed to improve AINo metadata opt-out available; in-app data can be opted out.
StandardAlways enabledContributed to improve AINo metadata opt-out available; in-app data can be opted out.
PremiumAlways enabledContributed to improve AINo metadata opt-out available; in-app data can be opted out.

The practical effect is a two-tier arrangement: the same product policy grants different withdrawal rights depending on the plan paid for. Organizations on lower tiers cannot decline the metadata contribution at all, even where their usage patterns reveal internal structure.

De-identification, removal, and retraining

When a customer opts out, or removes an application, the contributed data is removed on separate schedules. In-app data is removed from the datasets within 30 days; metadata is removed within 90 days. Models trained earlier on the contributed data are retrained afterward, rather than simply having future contributions stopped.

Two further safeguards are stated in the official FAQ. The company does not use data from applications that have no settings configured, and calls made to third-party large language model providers run under a zero data retention arrangement, meaning the external model provider does not retain the submitted content.

Why metadata is harder to anonymize than it looks

Stripping names and email addresses removes direct identifiers, but structural data can still be re-identifiable in combination. A small organization, a distinctive workflow, an unusual project mix, or a particular usage rhythm can map back to a specific company even without a name attached. Aggregation helps when datasets are large and mixed, but groups that stay small or distinct remain traceable.

This is the core compliance tension in tiered privacy. The same processing activity receives different rights depending on plan level, and metadata that resists full de-identification sits in the category that lower-tier customers cannot withdraw from. Procurement reviews should treat the published de-identification description as a starting point, not as proof that nothing identifying remains.

Five questions to ask a vendor before accepting AI data terms

  1. Exactly which data categories enter model improvement, and what remains strictly within the customer tenant?
  2. Which plan tiers hold opt-out rights for each category, and what cannot be withdrawn at any tier?
  3. What retention and removal timelines apply after opt-out or app deletion, and are previously trained models retrained?
  4. Do third-party model providers retain any submitted content, and under what retention commitment?
  5. How is de-identification verified, and what happens when small or distinctive datasets can be re-identified in aggregate?

Frequently asked questions

Can Free or Premium customers stop their metadata from being used for AI improvement?

Under the published policy, metadata contribution stays enabled on Free, Standard, and Premium plans with no opt-out. The in-app data contribution can still be declined by customers on those tiers; the metadata category is the one without a withdrawal option.

Does removing names and emails mean the data is anonymous?

Not necessarily. Direct identifiers are removed, but workflow structure, project patterns, and usage behavior can still point back to a specific organization, especially for small or distinctive deployments. Aggregation reduces this risk but does not eliminate it.

What happens to models already trained on my data after I opt out?

The stated process retrains the previously trained models after the contributed data is removed from the datasets. In-app data is removed within 30 days of opt-out or app deletion; metadata is removed within 90 days.

Atlassian Announces AI Training Data Collection—Free and Paid Users Can't Opt Out