From August 17, Jira and Confluence content feeds Atlassian's AI by default on cheap plans — and only Enterprise can refuse metadata collection at all.

On August 17, 2026, Atlassian starts using customer data from Jira, Confluence and its other cloud products to train and improve its AI. The policy went up in April. We are running it now because the deadline is close and the defaults are the story.

Two categories. Metadata is de-identified, aggregated signal: readability scores, complexity ratings, semantic similarity between pages, story points, task classifications, SLA values. In-app data is what you actually wrote — Confluence page titles and bodies, Jira issue titles, descriptions, comments, custom workflow names.

What you control depends on what you pay

From Atlassian's own documentation:

• Free and Standard: metadata always contributed, no toggle. In-app data on by default; an admin can switch it off.
• Premium: metadata always contributed, no toggle. In-app data off by default.
• Enterprise: metadata on by default, but can be turned off. In-app data off by default.

Read that again. The cheaper your plan, the more of your content ships by default — and only Enterprise can escape metadata collection at all. Your org inherits the defaults of its highest active plan, trials included.

Do this before August 17

Go to Atlassian Administration → Security → Data contribution and turn in-app data off. That is the sensitive half: ticket titles describing unshipped work, client names sitting in comments.

Be clear about what that does not fix. Metadata collection continues on Free, Standard and Premium regardless. There is no switch. Atlassian says data de-identified and aggregated at customer level may be retained up to seven years. If you opt out or delete an app, in-app data leaves the improvement datasets within 30 days and metadata within 90, and affected models get retrained.

Exempt entirely: customer-managed keys or BYOK, Atlassian Government Cloud, Atlassian Isolated Cloud, HIPAA-compliant orgs, and certain government and financial services customers.

Should you leave?

If your Jira describes unreleased product or named client specifics, this is a fair reason to reprice the tool. But read the alternative's terms with the same eye before you move. Linear, Shortcut and Notion all have their own AI training language. "Not Atlassian" is not a privacy policy.

Reported by Software Crit from the sources above. Every story is confirmed against at least two independent publishers before publication.