AI descriptions with EPFL
In early 2023, I had the opportunity to design an experimental feature for the Wikipedia Android app that brought machine-generated short descriptions to users. The goal: support editors—especially newcomers—in writing better short descriptions while evaluating the strengths and limitations of the AI model behind it.
This was a close collaboration between the Android team at the Wikimedia Foundation, researchers at EPFL—most notably Robert West and Marija Šakota—and our internal researcher Isaac Johnson, who helped guide the experiment design and evaluation.
Designing for human input first
The underlying model, Descartes, could generate high-quality short descriptions in over 25 languages. But the challenge wasn’t just technical—it was about how to integrate suggestions without diminishing the role of the human editor.
We intentionally kept machine suggestions hidden behind a tap. Editors had to actively choose to see them, then decide whether to accept, edit, or reject them. Suggestions appeared as tappable chips and could be edited before publishing. The interface encouraged thoughtful contributions rather than quick acceptance.
Onboarding showing machine suggestions
Key design principles
- Human in the loop: users remained in full control. The UI was built around guidance, not automation.
- Guardrails against harm: we introduced restrictions for biographies of living persons and added checks to minimize bias and hallucinated content.
- Encouraging learning: the onboarding experience explained what machine assistance was and prompted users to think critically before publishing.
Suggested descriptions as chips in UI
Results
The experiment ran from May 12 to June 16, 2023 and included thousands of edits across 25 language communities. Volunteer graders compared machine-assisted edits with human-written ones.
Suggested descriptions after tapping the chip
Some takeaways:
- Machine-generated edits scored almost as high as human edits (avg. 4.1 vs. 4.2 out of 5)
- 90% of accepted machine edits were graded 3 or higher
- Experienced users (50+ edits) produced higher-quality results and completed more edits
- Machine suggestions were modified in 38% of accepted cases—confirming that editors engaged with and improved the suggestions
Only 0.5% of users reported issues, and just 20 machine-assisted edits were reverted across all languages—well within safe limits.
Tapping the chip fills out the input field
Based on these insights, the model was migrated and made available to communities that expressed interest. A lighter version of the feature, showing only the most accurate suggestions, is now available in selected languages.
Preview before publishing
Final thoughts
This project wasn’t about replacing humans—it was about helping them. We designed a system where machine learning could quietly assist, but never take over. The editors stayed in control, and the design respected their judgment every step of the way.
Final step after adding the suggested description