Page MenuHomePhabricator

Image Suggestions uses AI-generated images from Commons when adding images on English Wikipedia
Open, Needs TriagePublic5 Estimated Story PointsBUG REPORT

Description

Steps to replicate the issue (include links if applicable):

  • Open the Wikipedia app
  • Navigate to an article on English Wikipedia (enwiki)
  • Start the flow to add an image to the article
  • Observe the suggested images provided by the app

What happens?:
The app suggests images from Wikimedia Commons without filtering out AI-generated images.
A user reported that they added an AI-generated image based on the app’s suggestion, assuming it was appropriate since it appeared in other language versions of the article.

What should have happened instead?:
AI-generated images should not be suggested when adding images on English Wikipedia, or at least should be filtered out when possible, given that enwiki policy (WP:AIIMAGE) generally discourages their use.

Software version (on Special:Version page; skip for WMF-hosted wikis like Wikipedia):
Prevent the app from suggesting AI-generated images (based on Commons categorization or metadata) when users are adding images to articles on English Wikipedia.

Potential solutions

Event Timeline

Seddon renamed this task from App suggests AI-generated images from Commons when adding images on English Wikipedia to Image Suggestions uses AI-generated images from Commons when adding images on English Wikipedia.Apr 7 2026, 2:53 PM

@mfossati @APizzata-WMF can this be easily done in the image suggestion pipeline?

This is something that needs to happen in the pipeline. Moving it to tracking for Growth, because it also affects what images are presented in the Growth Add-an-Image newcomer task.

@Ahoelzl TL;DR: not easy.

Besides a tag that my former team introduced on Commons UploadWizard, I'm not aware of other methods to mark uploaded content as AI-generated:

Screenshot 2026-04-13 at 13.33.19.png (1,868×2,030 px, 408 KB)

From a technical perspective, we could add a filter in the pipeline.
IIRC those tags live in the wikitext of Commons File pages, so parsing becomes a requirement. Moreover, this would be a partial solution, since the presence of such a tag depends on the uploader's discretion. Untagged AI-generated content can still be uploaded.
This is also a language chapter-specific requirement, since the suggestion came from another Wikipedia, where the image was allowed.

Looping in @Cparle, as I might have outdated information on AI-generated content handling on Commons.

IIRC those tags live in the wikitext of Commons File pages, so parsing becomes a requirement

Specifically, the tag adds the PD-algorithm template.

Looping in @Cparle, as I might have outdated information on AI-generated content handling on Commons.

I don't know anything more than what you're posted above ¯\_(ツ)_/¯

fwiw the image mentioned in the task also has a structured-data claim of "instanceof : AI-generated image" (Q116177346). Although this was added manually, could it be part of the pipeline for selecting or filtering images?

@GGoncalves-WMF who is the product owner of image suggestion generation logic? I think DPE has agreed to own the pipeline maintenance. In order to make this task actionable, someone has to do a lot of work to figure out the right thing to do for the product.

@GGoncalves-WMF who is the product owner of image suggestion generation logic? I think DPE has agreed to own the pipeline maintenance. In order to make this task actionable, someone has to do a lot of work to figure out the right thing to do for the product.

Originally the Growth team worked with the Structured Data/Content team to define image generation logic. Now that there is no Structured Content team, I imagine Growth's PM is the owner (hi, it's me!).

From the product perspective, I think it's likely best to just avoid suggesting AI-generated images across all wikis. Individual wikis can decide they are OK for certain articles, but for a cross-wiki tool, I think it's safest to assume we should avoid surfacing them as suggestions since newcomers are unlikely to know their local wiki rules on this topic.

If we agree with that direction, this solution seems like the most ideal one:

filter images with the {{PD-algorithm}} template in their Commons File pages

I realize, though, that this isn't a particularly easy change. So I’m curious whether this complaint has come up repeatedly, and how urgent we think it is. I know it came up once after a bad edit, but my impression was that this has been a relatively rare issue.

That said, now that we’re unblocked in scaling Add an Image and could potentially scale this task more widely, I wonder if this could become more visible and therefore more important to address.

Agree with Growth as product management owner.

In previous discussions on DE onboarding image suggestions, I think we were headed in the direction of Growth owning the product roadmap and requirements, with the pipeline itself still a bit fuzzier/co-owned between Growth and DE, to be further fleshed out when a product improvement came up. Since then we've started using the Service Catalog ownership model in other projects, and can probably try to formalize things in that framework.

All of that is to say: Kirsten, if this is a priority, we can partner on this.