Two conversations about artificial intelligence are happening on Wikimedia Commons at the same time, and they pull in opposite directions.
The first is defensive: how should Commons handle AI-generated images, and how can anyone tell them apart from photographs? The second is hopeful: could AI help people find the right file among more than a hundred million?
I follow both closely, as a contributor, a reviewer and a juror. This post is my attempt to set out where I think the real problems lie. It is a personal view, not a summary of any policy.
Why AI images are a Commons problem
Commons is an archive of evidence as much as an archive of pictures. When a Wikipedia article about a bird, a temple or a historical event shows an image, readers assume it depicts something that exists or existed. That assumption is the whole value of the archive.
An AI-generated image of a "traditional festival" or a "rare species" can look entirely plausible and depict nothing real. Placed in an article, it is worse than no image: it is a confident error.
Commons has a dedicated page on AI-generated media, and the community discussion around it keeps evolving. Some uses are clearly reasonable, such as an illustration of what AI image generation produces, clearly labelled. Others are clearly a problem, such as AI images presented as photographs of real places or people.
Identifying them is getting harder
As a juror and Quality Image reviewer, I spend a lot of time looking at photographs at full size. A few years ago, generated images often gave themselves away: impossible hands, melted text, repeating textures, lighting that came from nowhere. Those signs are fading.
What still helps, in my experience, is context rather than pixels:
Does the file page make sense? Real photographs usually come with a specific place, date and camera. Generated images often have vague descriptions.
Does the uploader's history make sense? A contributor with consistent photographs from one region is different from an account uploading polished images of subjects across several continents.
Is there metadata? Camera data can be stripped or faked, but its total absence on a supposed phone photograph is worth a second look.
Does it match other images of the same subject? A monument has one shape. A real bird species has real markings.
None of these is proof. Together they are a reason to ask questions before an image goes into an article.
Competitions have to decide this clearly. For the Wiki Science Competition 2025 in India, where I was National Lead, AI-generated images were not eligible. I think every documentary photo campaign should state the same rule explicitly and tell jurors how to handle suspected cases.
The hopeful side: finding things by meaning
The other conversation is about search, and here I am optimistic.
Commons search today depends heavily on file names, descriptions and categories. If the uploader called a file "IMG_2041" and categorised it poorly, a perfect photograph of exactly what you need may be invisible. That is one reason I spend so much time on categorisation.
Semantic search tools, such as WISE, try a different approach: they look at what an image depicts and match it to the meaning of your query, not only to its words. A teacher could search for "children learning in a village classroom" and find relevant images even if none of the files used those exact words.
I am very interested in this direction because it changes who can find what. It could surface excellent files from smaller communities whose descriptions are in languages a keyword search does not handle well. It could also help curators like me find unused images that belong in Wikipedia articles.
It also needs care. Semantic search should help people find real media. It should not quietly blur the line between photographs and generated images, and it should be transparent about why a result appears.
Making files more inviting
A smaller idea I followed was custom thumbnails for files, a proposal discussed through the Community Wishlist. The thinking, as I understood it, was simple: a well-chosen preview frame or thumbnail makes a file more inviting to click, especially for video, where the default frame is often a blurry first second. The proposal was eventually closed. I still think the underlying problem is real: discoverability is not only about search ranking but about whether a result looks worth opening.
Where I land
For me the two conversations meet in one principle. Commons is valuable because people can trust what it shows and find what they need. AI threatens the first and could greatly help the second.
So I want two things at once: clear rules and good habits for keeping generated images out of places where they mislead, and better, transparent tools for helping people find the real images that volunteers have spent years contributing. Both depend on the same unglamorous foundation: honest descriptions, accurate categories and contributors who care about context.