How do intelligent search, predictive collaboration, and workflow acceleration respect data privacy?
One of the best parts of working at Atlassian is getting to use our own tools, and we use them for all types of work. As a public company, we understand the importance of proper care for confidential information. Rest assured that we’ve thought deeply about how to prevent ML models from unintentionally revealing information about source data they’re trained on.
We’ve set up privacy controls both within your organization (to prevent people within an organization from seeing information they shouldn’t), and beyond (to prevent anything from getting outside your organization):
- We build experiences that respect the privacy controls in our products - eg. users will not be recommended content they do not have permission to see
- Where we build models that learn patterns from your data (e.g. from search queries that are made by your users on your instance), that data does not leave your group permissions for broader model training - de-identified search query strings are only accessed by automated jobs and are not read by individuals (restricted pages are never used)
- Where we build models that learn trends across customers (e.g. users generally search for things they recently worked on), we only source data from information like de-identified behavioral analytics (e.g. number of likes) and one-way vectorized content, aggregated across Cloud customers
- We use models trained on public datasets (i.e. which do not contain customer data), when possible
Example 1: we know privacy is important within your company, including Confluence page visibility settings. Intelligent search observes all group permissions settings, including user-level permissions, so Confluence pages marked “private” won’t be surfaced in recommended search results to users in your company without page access.
Example 2: we know you don’t want other companies to receive suggested search results based on your confidential information. Unlike other types of search engines, intelligent search does not aggregate top searches across customers to improve its functionality (it learns about individuals' search preferences), so machine learning models will always observe your group permissions and prevent information bleeding between customers. In other words, if you have a private Confluence space and your team is using it to collaborate on Confluence pages called “New Co acquisition”, other Atlassian customers using Confluence and searching “New Co” won’t see suggested results based on your “New Co acquisition” pages or users' search queries.