Designing AI-assisted tools a community could trust

Challenge

Wikimedia Commons needed millions of new metadata tags to make 140 million media files searchable — far more than any team could add by hand. The volunteer community was wary of automation, and any tool that felt like it overrode their judgment would fail.

Solution

Through interviews and community consultations, I learned that verification was what would earn trust, not automation. I designed a computer-vision-assisted tagging tool around one rule: suggest, never auto-apply. Every tag needed a human confirm, reject, or skip.

Outcome

Over 5,800 users adopted the tool across more than 340,000 files, with most usage coming from casual contributors tagging their own uploads. The community went on to build their own follow-up tools on top of the approach.