During Wikimania 2025, I had the opportunity to facilitate the first-ever AI table at Wikimania and share with Wikimedians how AI/ML is used in Wikipedia. Here is the slide deck I presented: https://docs.google.com/presentation/d/1aH-6VNxebo34cSxulyP1oYdzEUFiggVjNkaN0K7Nx0E/edit?usp=sharing

We discussed how the ML team trains models using Airflow pipelines, serves them via LiftWing, and the various features these models power on Wikipedia.

The conversations were very engaging, participants were open-minded and especially interested to learn that we primarily use predictive ML models, not generative AI. The live demos (1, 2, 3) showcasing ML models in action within Wikipedia features were particularly popular.

Below is feedback and ongoing questions:

1. Tone-Check Notice:
   • Some users found that the tone-check notice appears intermittently.
   • I also experienced this during the demos. Fortunately, the page where tone-check is enabled includes a note: Currently using experimental edit checks. For testing purposes only.
   • This helped clarify that the feature is still under development and will be more stable when deployed in production.

2. ArticleTopic and RRLA Model Documentation:
   • For articletopic, queries like articletopic:president Trump didn't work, while articletopic:chemistry mercury did. Participants wanted to know which topics are supported and suggested this information should be easy to find.
   • For the RRLA model, users wanted to easily find which languages are supported.
   • While this information is available in the model cards (article-topic model card, RRLA model card), it is not easily discoverable from the Wikipedia front-end.
   • Suggestion: Add links to model cards or documentation in the UI so users can easily access up-to-date information about topics and languages supported by the models.

3. RRLA Model Label Translation:
   • On ml.wiki, the "Likely to be reverted" filter label appears in English as shown in the screenshot below:
     
   • This label should be translated to Malayalam.
   • This translation should also work for other languages supported by the RRLA model.

4. Content Translation Tool:
   • Some users found the "issues" section in the content translation tool restrictive, as it prevents publishing until all issues are resolved. This leads some editors to bypass the tool, edit the article directly, and publish without using the content-translation tool altogether.
   • A user also asked why Google Translate is used in content translation, whether it is open source, and if Gemini (which they felt is better) could be used instead.

5. AI Proofreader Script:
   • After activating the AI Proofreader script in common.js, it worked as expected on en.wiki but not on ml.wiki.
   • We followed up with the script maintainer, and this was fixed as shown here: https://en.wikipedia.org/wiki/User_talk:Polygnotus/Scripts/AI_Proofreader#support_in_different_language

6. Wikimedia Commons Search:
   • When uploading an image taken in Nairobi National Park that has among other things a giraffe in the background, searching for "giraffe" in Commons did not return the image if "giraffe" was not in the description, caption, category, or title.
   • Suggestion: Can we use AI/ML in Commons search to match keywords with objects detected in images?

Thank you to everyone who participated and shared feedback. These insights are invaluable for improving our AI/ML-powered features and making them more accessible and useful for the Wikimedia community.

Thanks so much Mr. Bazira! I heard folks who stopped by appreciate the info you shared, and it looks like you got some good user testing done too ✨ @SSalgaonkar-WMF might be useful for you too ^_^