Jump to content

Commons:Requests for comment/Policy update for AI content

From Wikimedia Commons, the free media repository

Background

[edit]

Large numbers of images that do not reflect reality are being uploaded. When an inferior AI image is inserted into an article it disincentivizes actual photographers and graphic artists from creating images because they are more inclined to create an image for an article that has none. AI-altered images make the detection of violations of copyright more difficult. Our current policies were not written with this in mind. This update aims to improve our content policies to enable our contributors and visitors to easily tell when they are looking at AI content while still respecting the needs of individual projects who in some cases wish to allow AI-generated content.

This update requires that some AI content is marked as such in-image. This update does not otherwise editorialize projects outside of Commons. Commons has no intention of editorializing other projects, we only want to ensure we are not contributing to misinforming the public.

Changes to project scope policy

[edit]

1. A new section on Commons:Project scope about requirements for AI content:

The following kinds of AI generated files are not allowed on Commons unless it is indicated in-image and discernible in a 250 pixels wide thumbnail that the work is not authentic. One way to achieve that is by including the initialism "AI" in-image. A watermark from an AI company is not sufficient if it requires familiarity with the company to be understood as AI. AI-generated works that have been covered by professional press and AI-altered works (see below) for which the unmodified original has been uploaded are exempt from this requirement. For files uploaded before [$proposal_acceptance_date] there is a grace period until [$proposal_acceptance_date+1_year] to allow uploaders and projects to make their files compliant or upload them locally.

[list of types that get support from the voting sections below]

AI-generated/modified files that comply with these requirements still need to comply with all other relevant requirements (like having educational value) to be allowed on Commons.

2. Addition of a bullet point to "Files that are currently in use may still be subject to deletion for reasons beyond their scope" on Commons:Project scope which links to the new section:

Please vote on the subsections below. The AI content section and bullet point won't be added if no subsection manages to pass. (status quo) - Alexis Jazz ping plz 16:37, 23 June 2026 (UTC)[reply]

Votes and comments

[edit]

You may vote in all five subsections:

AI upscaling

[edit]

The AI-content section should include the following:

"Visual works that were upscaled using AI are not acceptable unless the inauthenticity is indicated as described above and the file page is tagged with {{AI upscaled}}."
AND - support applying THE SAME STANDARD to ALL such modified files (not just ai!). Lx 121 (talk) 14:43, 26 June 2026 (UTC)[reply]

AI-generated photorealistic images and videos

[edit]

The AI content section should include the following:

"AI-generated works where a reasonable person may perceive the work, not considering the filename or description, as a video or photo taken with an actual camera are not acceptable unless the inauthenticity is indicated as described above."
Over 1.8 million views since 2023 without watermark - how much harm has this really done?
Over 700,000 views since 2019
  •  Oppose I'd support a mandatory disclosure of inauthenticity in the file description for all images that are likely to be mistaken for authentic photos - something that Commons policy does not require yet for any images, even deliberately misleading ones (see below). However, this ill-prepared RfC fails to consider several aspects that make the particular proposal here a bad idea:
    • As others have noted, the proposed permanent watermarking scheme is intrusive (deliberately so, given the 250px requirement) and very destructive (note that Alexis Jazz' proposal doesn't even allow for preserving the original image; is there any precedent for such a policy on Commons?).
    • As such, a strong rationale is needed for such extreme measures. And in the very non-neutrally worded intro to this RfC, Alexis Jazz fails to provide such a strong justification, instead making several unsubstantiated and IMO factually dubious claims. For example that Large numbers of images that do not reflect reality are being uploaded (for which, as pointed out below, no evidence had been given): Assuming that this doesn't just mean artworks or other non-photo illustrations (in which case it is of course true), but AI-generated images of the kind that are the subject of this RfC: How many of these are really being uploaded and can't be deleted under existing policies? Any concrete examples that are actually problematic in the sense of the risks claimed by the RfC?
    • Since Alexis Jazz failed to provide such examples of problematic INUSE images, I'm including two examples of photorealistic AI-generated images here that definitely match the "a reasonable person may perceive ..." definition and have already been widely viewed since having been uploaded to Commons in 2019 and 2023, respectively. Can someone point to INUSEs of these two (i.e. a page on another Wikimedia project that uses them) that are likely to mislead and where this watermark proposal would help? I'm not aware of any. (Yes, at least one of them has been covered by professional press in the past, but I don't see how that meaningfully reduces the risk of misinformation and the other purported dangers claimed in the RfC.)
    • This update does not otherwise editorialize projects outside of Commons [my bolding] - but it very much editorializes them regarding such images. Basically, it looks like this RfC rests on Alexis Jazz' fear that other Wikimedia projects will use them in a way that spreads misinformation or otherwise furthers the RfC's claimed risks (by failing to include the kind of notice that Alexis Jazz wants to add with the watermark instead). One would think that we should have a little more evidence that this is happening and is an actual problem, before embarking on such a serious deviation from COM:INUSE and constructing ways for Commons to inject its own messaging into local projects' use of media from Commons.
    • And on the other hand, we have strong evidence that the risks from non-watermarked inauthentic images may be much lower than the RfC presumes. Namely, manipulated photos and other artificial imagery that a reasonable person may perceive [...], not considering the filename or description, as a video or photo taken with an actual camera go back to the 19th century, see e.g. en:Photograph manipulation, and we have many such images on Commons, see e.g. Category:Altered Soviet photographs or the examples below. Does Alexis Jazz also feel the need to improve our content policies to enable our contributors and visitors to easily tell when they are looking at such manipulated or artificial images that are not AI-generated? Again, we don't even have a policy yet that requires them to be clearly described as inauthentic in the file description, which is at odds with assumptions that this is a major problem.
Not an authentic photo of the actual Benjamin Franklin. From en:Darren_Aronofsky#Nonfiction_work: However, Deadline suggested the series may be 'ahead of its time' and that Aronofsky could ultimately be remembered as a pioneer in AI-driven storytelling.
    • And, aside from deliberate manipulations/photoshopping that is intended to mislead, there is another whole category of images that do not reflect reality but are made to look like authentic photos or videos, namely stills from (notable) movies or TV shows depicting actual historical events. We have always permitted these on Commons too when COM:L allows it, and they are INUSE in many Wikipedia articles (without burnt-in watermark warnings like "FROM A FICTITIOUS MOVIE"). With Hollywood likely to use a lot more generative AI in the near future, this points to another aspect in which this proposed rule and its commingling of "AI-generated" and "inauthentic" has not been thought through. The image on the right is an early example of the kind of valuable AI-generated film still uploads we can expect, also for more recent historical topics (valuable not as illustrations for a Wikipedia article about the depicted person or historical event of course, but about the film itself).
    • To point out just another indication that this proposal was not fully thought through: I don't see any consideration of Commons:Language policy in the RfC (even when restricting the watermark to just "AI" instead of a more informative "AI-generated" or such, that's not language-neutral, as Alexis Jazz concedes below; e.g. it's "IA" in French. And the mandatory deletion of the non-watermarked original will make it difficult to impossible to produce a version with a localized watermark later).
Regards, HaeB (talk) 07:40, 25 June 2026 (UTC) Edited to correct the attribution of the comment linked in the second bullet point. HaeB (talk) 06:00, 26 June 2026 (UTC) [reply]

AI-generated artworks

[edit]

The AI content section should include the following:

"AI-generated works where a reasonable person may incorrectly perceive the work, not considering the filename or description, as a historical or commercial artwork or an artwork by a known artist are not acceptable unless the inauthenticity is indicated as described above."
AND - support applying THE SAME STANDARD to ALL such modified files (not just ai!). Lx 121 (talk) 14:49, 26 June 2026 (UTC)[reply]

AI-altered visual works

[edit]

The AI content section should include the following:

"Visual works that were modified using AI tools (e.g. restoration) are not acceptable unless the unmodified original is also uploaded and linked from the file page or available in the file history of the modified work. If this is not possible the inauthenticity should be indicated as described above."
AND - support applying THE SAME STANDARD to ALL such modified files (not just ai!). Lx 121 (talk) 14:54, 26 June 2026 (UTC)[reply]

AI-generated perceived authoritative representations

[edit]

The AI content section should include the following:

"AI-generated works where a reasonable person may perceive the work, not considering the filename or description, as an authoritative representation of a subject are not acceptable unless the inauthenticity is indicated as described above."

 Support Nettime Sujata (talk) 02:43, 24 June 2026 (UTC)[reply]

ONCE the intended reach has been clarified: oppose watermarking as the "solution", this is overkill , support including info in the file description (& possibly adding it to the file metadata)?.
AND - support applying THE SAME STANDARD to ALL such files (not just ai!). Lx 121 (talk) 15:02, 26 June 2026 (UTC)[reply]

Discussion

[edit]

Some questions you might have:

What is a "reasonable person"?
A reasonable person is a legal fiction. In this case, it is a person who is not a subject matter expert, not particularly inept but also not quite as capable as an average Joe. Some people have developed a "sixth sense" for even minor flaws in AI content, but this is an exception. An elderly person of sound mind is also a reasonable person, and many of them have great difficulty telling plausible AI-generated photorealistic images from actual photos.
What does "A watermark from an AI company is not sufficient if it requires familiarity with the company to be understood as AI." mean?
It means that a four-pointed star in the bottom right corner or a cloud with four-pointed star pupils ain't gonna cut it.
What is an "AI-generated perceived authoritative representation"?
This was added in response to this comment by Ckoerner: That Cleopatra image is exactly why I'm concerned about AI media on Commons. It's the default image used when this article is shared on social media. Out of context, does it over time become the default representation of the subject? Preferred over other art representing the subject? Used because it's the most "interesting"? The image has no basis in reality. There's no research into what dress folks wore back then, no careful representation of ethnicity or culture, it perpetuates modern beauty standards, etc. It's a refinement of noise from a sludge of data.
Will the grace period information be in the policy forever?
Some time after the grace period ends it'll be historical information and it may be archived.
This isn't fully banning AI upscaling!
It requires in-image clarification and file page tagging, making the files highly visible. This gives us the tools to make maintenance categories and ask the uploaders to upload the unmodified original and/or track the original down ourselves. If such an upscaled file goes unused it can be deleted as a non-notable artwork.
This isn't fully banning AI slop!
If it's unused, it can generally be deleted as a non-notable artwork which is out of Commons' project scope. That will be a community decision to be decided on a case-by-case basis.
Do we now have to clarify the AI nature of File:AI generated woman with extra legs.jpg in-image?
A reasonable person would not perceive that image as a photo taken with an actual camera.
Why do indications of AI need to be in-image, can't we just put that on the file page?
When a thumbnail is inserted on a project a visitor may never see the file page. They will see the thumbnail, and if they click it, they will see a larger image in Media Viewer. When an image is hotlinked outside of Wikimedia (which is allowed), the file page may or may not be linked at all.
This will push projects to upload their files locally!
That's doubtful. Hosting content that isn't super educational doesn't really hurt our mission, but misinformation does. If any individual projects want to mislead their readers they should host their own slop instead of putting that burden on the Commons. - Alexis Jazz ping plz 16:43, 23 June 2026 (UTC)[reply]
I'm sympathetic with this essay. Sapphaline (talk) 17:19, 23 June 2026 (UTC)[reply]
I think all the suggestions are good. But they should only apply to files they are in use. All not used AI generated content should just be considered as out of scope. I would suggest the following policy: All AI generated content is considered as out of scope and may be deleted under the "F10. Personal files by non-contributors" speed deletion criteria unless one of the following requirement is met: 1. The work was published in a peer-reviewed scientific journal. 2. The work is subject of non-trivial coverage by reliable sources or part of a set with this coverage. GPSLeo (talk) 17:39, 23 June 2026 (UTC)[reply]
"The work was published in a peer-reviewed scientific journal" - that doesn't guarantee anything. Sapphaline (talk) 17:43, 23 June 2026 (UTC)[reply]

Should watermarking be reversible?

[edit]

Watermarking is a destructive process. It irreversibly modify pixels overlapping with and around the watermark. Should we set out methods to add watermarks in a nondestructive or reversible way? Approaches I can think of:

  1. Add a machine-readable metadata flag to the description page, then add a pre-processor in our thumbnail rendering engine to add the watermark to show them on wikis that use them (impractical due how how "quick" WMF implement things, 02:36, 24 June 2026 (UTC));
  2. Encourage users to upload the original version, and a bot will add the watermark afterwards, keeping the original copy in the file history.

1F616EMO (talk) 17:52, 23 June 2026 (UTC)[reply]

Of course not. If a reader doesn't want to waste their time on AI slop (= any image generated by AI), we should respect their choice, and watermarking the initial image is the only reliable way to do this. Sapphaline (talk) 17:54, 23 June 2026 (UTC)[reply]
I am not oppose to watermarking; in fact, this thread based on the assumption that watermarking will become policy. The question is: while the right of rejecting AI slops is the default, should we respect their choice of opting-into AI-generated materials if they explicitly ask for that? 1F616EMO (talk) 18:02, 23 June 2026 (UTC)[reply]
1F616EMO, pre-processor in our thumbnail rendering engine
Believe it or not, exactly this (including the machine-readable metadata) had crossed my mind while drafting this RfC. However, I'm also a contributor to Phabricator and familiar with typical development timelines. If prioritized, such a feature could be implemented in -maybe- 3 months. But this is unlikely to get prioritized. Which means it would likely take (and let me stress this is not an exaggeration) more than a decade, if it happens at all.
Encourage users to upload the original version, and a bot will add the watermark afterwards, keeping the original copy in the file history.
Similar to what DatBot does on English Wikipedia to enforce their fair use policy, a similar bot could be created to aid the enforcement of watermarking. English Wikipedia's NFC policy does not depend on DatBot however, and neither does this RfC. Tools would certainly be helpful, but are not a prerequisite. - Alexis Jazz ping plz 20:57, 23 June 2026 (UTC)[reply]
Then (2) will be our only practical approach. The problem is that the bot have to find ways to prevent itself from going into an infinite loop, or fight with other bots that also do minor adjustments. 1F616EMO (talk) 02:38, 24 June 2026 (UTC)[reply]
That could be solved in various ways. An obvious one would be for the bot to also categorize the image as having the required watermark when it overwrites the image. - Alexis Jazz ping plz 03:08, 24 June 2026 (UTC)[reply]
IMO, applying a "destructive process" to slop is a good thing, for upscaling and restoration support option 2. ⇖ /.°°.\ ⇗ (They/Them/Their) 14:21, 25 June 2026 (UTC)[reply]
these should be tagged in the description, and a new software feature added to add a language-specific AI watermark over the top of the image, that doesn't affect the image itself. This may take a while to implement, but not as long as the policy's grace period.
@Yodin: are you familiar with typical developer timelines? For example, w:WP:THEYCANTHEARYOU was like a running gag for many years. This idea may not be as simple as you think it is. You'd probably need a way to customize where the watermark appears. You may need different watermarks for different situations. SVG and video will be special cases. Thumbnail sizes were recently standardized to reduce cache storage, but caching would explode for language-dependent watermarks. And all this has to play nice with third party libraries that MediaWiki uses. - Alexis Jazz ping plz 02:57, 26 June 2026 (UTC)[reply]
The pros of this approach massively outweigh the cons. As well as supporting different languages, it would provide a standardised symbol which would make life much easier and less confusing for readers, and (as you say) could allow several different types of watermark depending on what exactly AI has been used for (upscaling, full image generation, alterations, etc.). There are also several different technical approaches that could be used to implement this, many of which would be much easier than you suggest. For example, it could be a separate SVG image automatically superimposed over the original image, which would also allow images to be resized to any dimension and still have the watermark clearly visible, and wouldn't affect caching. I think any technical approach like this would be worth waiting for. ‑‑YodinT 08:44, 26 June 2026 (UTC)[reply]
I agree that this approach is preferred, but we can't wait indefinitely. 1F616EMO (talk) 08:47, 26 June 2026 (UTC)[reply]
The current proposal says it can wait for one year. I think that the damage caused by the proposal (irreversible watermarks, deletions of many valuable works along with the slop, etc.) outweigh any temporary benefit, and would also almost certainly prevent any development of a better technical solution. Your suggestion of having users upload an unwatermarked image, and a bot adding the watermark would also be much better than the proposed solution, which seems to be being rushed through without thinking of the consequences. ‑‑YodinT 12:30, 26 June 2026 (UTC)[reply]
Well, it would likely take (and let me stress this is not an exaggeration) more than a decade, if it happens at all. Can we really wait that long? 1F616EMO (talk) 04:55, 27 June 2026 (UTC)[reply]
This proposal is full of assertions without any supporting evidence, and that estimate was for a much more complicated solution (rewriting the thumbnail renderer to add a watermark into every AI image). ‑‑YodinT 08:34, 27 June 2026 (UTC)[reply]
Why can't we wait for a technical solution (of superimposing labels in a systematic way without changing the file content)? If the answer is because we will be inundated before a technical solution, I would like to see data for that. How many AI-generated images will be there on Commons in 1 year, in 2 years, based on the last 2 years' data? whym (talk) 13:49, 26 June 2026 (UTC)[reply]
@Yodin: Another reason that supports option (2) is that the default download option will include the "AI" watermark. This aligns with the proposal's objective to (by default) clearly indicate the inauthenticity and extends it to offline usages. The original copy in file history remains an option for those who want to download the watermark-free version; clear instructions should be provided in the file description pointing to the original version. 1F616EMO (talk) 04:55, 27 June 2026 (UTC)[reply]

About watermark symbol

[edit]

Why it has to be "AI"? What if the initialism is different in my language? Shouldn't be more elegant to add the AI spark symbol as a watermark? Tmagc (talk) 19:33, 23 June 2026 (UTC)[reply]

  • The AI spark symbol originates from my box of detergent. (seriously, check the packaging of your detergent - it's likely on there!)
    Tmagc, it doesn't have to be, but it is suggested as an option that is always considered valid. I've chosen "AI" because it is understood (according to Wiktionary) in English, Danish, Dutch, Esperanto, Indonesian, Italian, Japanese, Malay, Maltese, Swedish and Vietnamese. The inverted "IA" (speakers of languages who typically use this likely also recognize "AI") is understood in Albanian, Aragonese, Catalan, French, Galician, Portuguese, Romanian and Spanish. Even without considering that English is very popular as a second language, "AI" can be understood by a large portion of the world population.
    What needs to be indicated is inauthenticity. Meaning that what File:Alex Pretti AI comparison.jpg does (and I checked - it's discernible in a 250px thumbnail) is a valid way to go about it, even though it doesn't specify AI, but it clarifies that the right side does not come from a camera. There is no specific language requirement in the RfC text. If an image clearly targets a German-speaking audience (for example if it includes German-language labels within the image) it would be appropriate to use "KI". - Alexis Jazz ping plz 21:21, 23 June 2026 (UTC)[reply]
    Just to add, in french many brand communication directly write AI, and not IA. Even when spoken, the advertisement are calling themselves AI (like at the end of this ad for Microsoft AI I believe it's the same in many other country.) Betelplouf (talk) 08:32, 24 June 2026 (UTC)[reply]
  • The proposal says "One way to achieve that is by including the initialism "AI" in-image." [my italics]. So, other ways can be used too.
    Perhaps this kind of misinterpretation would be less likely if more than one example was provided. -- Jtneill - Talk 10:29, 24 June 2026 (UTC)[reply]

AI-generated diagrammatic images

[edit]

I've seen a substantial number of AI-generated images which would not be covered by any part of this proposal, generally skewing towards diagrammatic content:

  • AI-generated infographics (e.g. ones created by NotebookLM)
  • AI-generated maps
  • AI-generated graphs, flowcharts, and other diagrams

Personally, I would prefer that all of these be disallowed from upload entirely, as they frequently contain major inaccuracies which are impractical or even impossible to correct - but I'm sure others feel differently. Any chance this proposal can be expanded to encompass them? Omphalographer (talk) 21:59, 23 June 2026 (UTC)[reply]

  • Omphalographer, this is a solid suggestion for a future RfC. Making an RfC too broad makes it more likely to fail, so in this one the focus is on ensuring everyone is informed. We almost certainly will need more RfCs to deal with all the challenges AI presents. Drafting a solid RfC often takes weeks if not months, if I make it look easy my apologies. - Alexis Jazz ping plz 23:04, 23 June 2026 (UTC)[reply]
I think one potential option would be simply categorizing informatic AI images as LLM contributions, which are already basically banned under WP:NOLLM Mausfield (talk) 00:36, 24 June 2026 (UTC)[reply]
w:WP:NOLLM is an enwiki policy and does not apply to Commons. Omphalographer (talk) 00:42, 24 June 2026 (UTC)[reply]
& many of them are perfectly valid & accurate, & the accuracy rate on properly prepared & prompted ai "diagrammatic images" is IMPROVING faster than the error-rate on human-created ones. a "blanket ban", as you seek, would be MASSIVE over-reach.
IF you see a file with an error, fix it, talk-page it, or DR it (as you have done so often, whether the file contained errors or not). that is how Commons works. Lx 121 (talk) 15:17, 26 June 2026 (UTC)[reply]

Actual proo

[edit]
  • Large numbers of images that do not reflect reality are being uploadedTemplate:Cn
  • When an inferior AI image is inserted into an article it disincentivizes actual photographers and graphic artists from creating images because they are more inclined to create an image for an article that has noneTemplate:Cn
  • AI-altered images make the detection of violations of copyright more difficult. Template:Cn

So is there any shred of proof of these things actually being the case? Or can we just create whichever fictional scenarios we want for the sake of policy pushing? --Trade (talk) 21:45, 23 June 2026 (UTC)[reply]

What does Template:Cn stand for? Tmagc (talk) 21:47, 23 June 2026 (UTC)[reply]
@Tmagc: citation needed, I'd assume. JJPMaster (she/they) 22:00, 23 June 2026 (UTC)[reply]
Tmagc, I'd also guess it's meant to refer to "citation needed".
It's not hard to find examples.For more, try Category:AI-generation related deletion requests. Note that not every AI-generation related deletion request gets categorized, deletion tags without discussion (e.g. {{No permission since}} and {{Copyvio}}) don't show up there and we don't always manage to actually detect AI-generated hallucinations. Still, the /deleted subcategory currently contains 1,637 pages. That should be sufficient reading material. - Alexis Jazz ping plz 22:26, 23 June 2026 (UTC)[reply]

Current exceptions in Wikipedia

[edit]

The English language Wikipedia holds an "Image use policy" that permits usage of certain AI-generated images: "Obvious exceptions include articles about AI and articles about notable AI-generated images [...]". This goes for most Wikipedia's, as far as I know. Does this mean that images used in articles like AI art are permitted as exceptions (after addition of the new AI-watermark)? Vysotsky (talk) 20:32, 23 June 2026 (UTC)[reply]

Vysotsky, yes. This RfC does not ban anything if it has a watermark. Images may be deleted on other grounds (copyvio, vandalism, etc) but COM:INUSE would protect them in most cases. The only way this RfC would cause a file to be deleted is if nobody is willing to add an in-image indication of inauthenticity. For the notable AI-generated images there's an exception, they do not need to be watermarked. Of course, any project community is free to upload a watermarked version of notable AI-generated images if they'd prefer to use that. - Alexis Jazz ping plz 21:48, 23 June 2026 (UTC)[reply]
NOW you're adding "special exceptions"? for "notable AI-generated images" - how exactly do we determine that? notable TO WHOM? are you TRYING to make discussions on here longer on purpose? xD !/? Lx 121 (talk) 15:30, 26 June 2026 (UTC)[reply]
tl,dr - i oppose watermarking, but IF it's the policy, then it should apply to ALL files in the applicable category. ONCE you start making "special exceptions", the whole concept fails; in both logical rationale & practical application. Lx 121 (talk) 15:30, 26 June 2026 (UTC)[reply]
Notable AI images should definitely stay, such as the one that won a contest and got into a controversy (forgor the name). It makes sense for it to be kept due to the educational value, though there have been people like @Danÿa pushing for a no exceptions policy. I would like to hear their full take. Dabmasterars [EN/RU] (talk/uploads) 20:54, 27 June 2026 (UTC)[reply]
I’m ok about accepting AI-generating stuff in the cases you mentionned, because they are notable for a reason or another. But besides these few cases, I’d argue for a total ban on AI-generated content for reasons I already mentionned in another comment. Danÿa (talk) 20:25, 28 June 2026 (UTC)[reply]
  • Question coming from Wikipedia. The Article w:Dead internet theory uses several AI images under the noted exception. Would any of these need a watermark?
GeogSage (talk) 05:08, 28 June 2026 (UTC)[reply]
GeogSage,From what I can tell the other images on that article are not AI. There is considerable overlap between images that could be affected by this RfC and w:en:Wikipedia:Image use policy#AI-generated images. - Alexis Jazz ping plz 17:48, 28 June 2026 (UTC)[reply]

We do already have guidelines that address nearly all of these in more detail; is this also a proposal to deprecate all of those?

[edit]

I haven't fully considered all of the proposals here yet, but my first impression is that we've spent an awful lot of time debating the finer points of multiple existing guidelines on these very same issues already. In some ways, these proposals are stricter; in some ways, looser. Since you're proposing modifying a policy, and given the relationship between policies and guidelines, isn't this also, practically speaking, a proposal to deprecate those guidelines (or substantially rewrite them)?

— Rhododendrites talk |  22:42, 23 June 2026 (UTC)[reply]

@Rhododendrites: this is a complicated topic and debate about it probably isn't going away anytime soon. My take on this: these are, as you say, guidelines. Guidelines are primarily that: guidelines. They suggest best practices, but are not necessarily enforceable. There are some ifs and buts to that, for example, consistently ignoring a guideline may in some cases be seen as disruptive, especially if there's community consensus to follow a guideline for a particular set of files. Guidelines may also (again, my personal take) clarify specific situations on which policy may be unclear or not sufficiently detailed. COM:PEOPLE could be an example of that, which is also referred to by policy. But guidelines are not policy, and definitely can't override policy. Commons:AI images of identifiable people has been debated a lot recently. It's written as if it's policy, but it's not. The truth is, it's a guideline we never really needed. It hasn't actually enabled the deletion of anything, despite getting cited as the reason for deletion in some cases. Unused non-notable AI-generated images of people can be deleted because they're not educational. If an image can be classified as defamation it can be deleted on those grounds, AI-generated or not. The guideline can't do anything about material that is in use because that's policy. And AI-generated images that don't show people, no matter how problematic, are not covered by that guideline. In my view, there's no difference between an AI-generated image of a celebrity or an AI-generated image of Lake Tahoe. Neither would be an accurate representation of the subject, so it's likely neither would be educational. The only difference is that a lake can't have its feelings hurt. But hurt feelings are not related to AI-generation, any real photo that isn't too flattering carries the same problem. And we regularly see people or someone representing them uploading shiny press photos to replace a freely licensed photo on their article. Or nominate the freely licensed photo for deletion because they don't find it a very flattering image.
This RfC presents a requirement to indicate inauthenticity in-image for various cases. If there's a guideline that says that isn't required, this will override that. In COM:AIIP there's a section that suggests it is sufficient to link an external source that hosts an unmodified image. Given the prevalence of link rot I consciously decided not to include that (instead requiring unmodified originals to be uploaded here for those cases), so if this RfC becomes policy that's a part of COM:AIIP that can be scrapped. AIIP also allows an AI-generated photorealistic image of a person if it was published by the person it depicts, even if the image hasn't been covered by the press. If his RfC becomes policy, the inauthenticity would need to be indicated in-image.
Regarding Commons:AI-generated media and Commons:Upscaling, it doesn't look like this RfC really clashes with them, but if you believe it does please elaborate. - Alexis Jazz ping plz 00:56, 24 June 2026 (UTC)[reply]

Clarification needed

[edit]

It would be helpful if these terms used above can be clarified:

  • "covered by professional press" - what is considered "professional press"?
  • "perceive the work...as a...commercial artwork" - what is considered a "commercial artwork"?
  • "authoritative representation of a subject" - this was mentioned in the Q&A, but the answer given is not really an answer. What is actually "authoritative representation of a subject", does it means images of people, or does it also include "non-photographic images like logos and emblems", as mentioned by Omphalographer here?

Thanks. Tvpuppy (talk) 01:55, 24 June 2026 (UTC)[reply]

Re. "authoritative representation" - and if it does mean images of people, this seems like it overlaps heavily with "photorealistic images". Omphalographer (talk) 02:49, 24 June 2026 (UTC)[reply]
  • There has been debate about this, and several Commons users don't seem to like having dependencies on definitions from another project (like Wikipedia's "reliable sources") in Commons guidelines or policy. They want Commons policy to stand on its own. Without a fixed list of organizations there will always be a grey area, that's unavoidable. If this actually ends up leading to regular heated debates on deletion requests we may create a guideline to define this term more tightly.
  • Commercial artwork: see Commons:Deletion requests/File:1960's art of cow getting abducted by UFO in midwest (ChatGPT Plus June 2026).png for an example.
  • That would include unofficial logos. AI-generated or not, those are generally out of scope and we would likely delete them anyway, assuming WMF legal doesn't beat us to the punch.
- Alexis Jazz ping plz 04:08, 24 June 2026 (UTC)[reply]
Thanks for your reply.
  • For the first point, I am still confused about the use of the uncommon term "professional press". It doesn't mean anything if it is not defined, because it is as confusing as if the proposal says the contrary: it applies to AI images covered by "unprofessional press".
  • For the second point, the term "commercial artwork" is misleading, I think it meant to say "physical artwork" or "human-made artwork" instead. It seems it's more about the artwork being incorrectly perceived to have physically existed in real life or have been created by a human, and less about the artwork being "commercial".
  • For the third point, I assume this meant that this section excludes AI images that are actually the "authoritative representation of a subject", for example a company's official logo that is AI-generated.
Thanks. Tvpuppy (talk) 05:28, 24 June 2026 (UTC)[reply]
[edit]

In the proposed text, I suggest linking "AI generated" to w:Generative AI instead of w:AI art because the proposed list of image types is broader than art (e.g., includes photorealistic images). -- Jtneill - Talk 03:58, 24 June 2026 (UTC)[reply]

I also suggest "AI-generated" instead of "AI generated" -- Jtneill - Talk 05:01, 24 June 2026 (UTC)[reply]

AI colorizations

[edit]

How about vintage/old images that were colorized using AI? Huge examples are most of the uploads of the notorious user Liane777 (talk · contribs). While colorization helps in generating colors for the convenience of several modern-day users, it only generates colors subjectively. There is a possibility of inaccuracy or lack of objectivity in the AI-generated colorizations. JWilz12345 (Talk|Contributions) 04:04, 24 June 2026 (UTC)[reply]

I think this is covered by the #AI-altered visual works section. Thanks. Tvpuppy (talk) 06:21, 24 June 2026 (UTC)[reply]
JWilz12345, in addition to what Tvpuppy said you may refer to Commons:Colorization which I wrote nearly 8 years ago. (I feel so old now.. ) - Alexis Jazz ping plz 02:07, 26 June 2026 (UTC)[reply]
colorization is colorization. whether it was done by ai, "older" digital editing (which is ALSO ai), hand painting, or whatever method. what technique was used is a matter for file info. it matters more exactly what changes were made & why (colours, choices of same, etc.). & NONE OF THEM are a physics-precise reconstruction of EXACTLY what the colours in the original scene were; they are ALL "guesses"/selections/choices of what colours to use; educated or otherwise... Lx 121 (talk) 17:21, 26 June 2026 (UTC)[reply]

Redundant text

[edit]

In the proposed text, the statement that "AI-generated/modified files that comply with these requirements still need to comply with all other relevant requirements (like having educational value) to be allowed on Commons." is not necessary (redundant) within the broader context of Commons:Project scope. -- Jtneill - Talk 04:07, 24 June 2026 (UTC)[reply]

Placement of proposed text

[edit]

Where is the proposed text to be placed in Commons:Project scope? Could a full draft of the proposed changed policy be provided? -- Jtneill - Talk 04:10, 24 June 2026 (UTC)[reply]

Perhaps this will be more suitable to be placed in a subpage, just like Commons:Project scope#Scope part 2: Pages, galleries and categories, given how complex and long this can potentially be. Thanks. Tvpuppy (talk) 06:20, 24 June 2026 (UTC)[reply]

AI Agent jailbreak

[edit]

What about the images that were created by gaslighting the AI Agent into forgetting its rules? I am not referring specifically to illegal content, but to content that violates the Terms of Service of the AI model's company (and such a violation may also constitute a civil wrong). Vellion Eagle (talk) 06:31, 24 June 2026 (UTC)[reply]

I'd say that would be considered the same as photographing in a museum with "photography forbidden" rules when entering (COM:HOUSERULES). I.e. we don't care about that. --Jonatan Svensson Glad (talk) 07:13, 24 June 2026 (UTC)[reply]
In some scenarios, this might represent irresponsible disclosure, and if the context was corrupted by the gaslighting, image artifacts might reveal the method used. And the vulnerability might not be patched. Vellion Eagle (talk) 07:53, 24 June 2026 (UTC)[reply]
It's still not Commons' role to secure infrastructure that isn't part of Commons' remit, or to mandate particular methods of contributing. Instead, policy has to be set at the levels of the user account (to counter serial disruption) and of the individual file (which in theory is performing a role at one or more wikis). Arlo James Barnes 21:13, 24 June 2026 (UTC)[reply]
An AI image is an AI image, and so has all of the drawbacks and negatives, and it doesn't matter whether it was generated with a prompt designed to evade input filters (I refuse to call this "prompt injection") or some other prompt, or whether it was generated with an uncensored or regular model. Also, the only possible use case for this in context of Commons I can imagine would be AI-generated images of various surgeries/injuries, something we absolutely should not allow here. Sapphaline (talk) 07:57, 24 June 2026 (UTC)[reply]

Wording

[edit]

For non-native english speaker terse wording here can be somewhat confusing. Proposals should have a bit more about the background before, such as "currently there is no restriction for AI-modified works, this section aims to limit or clarify their use and proposed wording .. etc.". Ipr1 (talk) 09:36, 24 June 2026 (UTC)[reply]

What about other media types such as audio?

[edit]

The proposals are limited to visual works and don't include other media such as audio. Consider rewriting to more inclusive of different media types.

AI upscaling limited to visual works

The proposed text for AI upscaling refers to "visual works" but not also to audio or other file types which can also be upscaled. I recommend modifying to be more inclusive of media types which can be upscaled using AI.

"Visual works that were upscaled using AI are not acceptable unless the inauthenticity is indicated as described above and the file page is tagged with {{AI upscaled}}."
AI-generated photorealistic images and videos

The proposed txt for AI-generated photorealistic images and videos doesn't include other media types such as audio. I recommend being more inclusive of different media types.

"AI-generated works where a reasonable person may perceive the work, not considering the filename or description, as a video or photo taken with an actual camera are not acceptable unless the inauthenticity is indicated as described above."

... and so on (I haven't l listed the other proposals but they have similar limitations)

-- Jtneill - Talk 09:44, 24 June 2026 (UTC)[reply]

I can't speak for the proposer, but audio files make up a small percent of files on Commons. See Special:MediaStatistics: as I write this, audio files are only 3.02% of the total file count and 0.237% of the storage. If we start seeing a lot of audio slop being uploaded we might need to revisit, but right now audio is not where the crisis is.
Does Commons even have any AI-upscaled audio files? Category:Media modified by AI doesn't have a subcategory for audio. Apocheir (talk) 00:40, 26 June 2026 (UTC)[reply]
Exactly. We may well need to look at audio in the future, but this RfC was already quite complicated. Audio requires a very different kind of approach and has different considerations. - Alexis Jazz ping plz 23:22, 27 June 2026 (UTC)[reply]

Scope or beyond scope?

[edit]

Point 1 ("A new section on Commons:Project scope...") seems to say that the proposed conditions for AI-generated content are about scope. (In other words, violating content would be out of scope.) Point 2 seems to say they are not about scope, because it categorizes them as one of "reasons beyond their scope". (In other words, violating content would have non-scope problems.) Isn't that a contradiction? --whym (talk) 09:51, 24 June 2026 (UTC)[reply]

No, it is a two step check. First check for the AI policy and only if the photo is fine under this policy the check about general scope has to be done. GPSLeo (talk) 15:42, 24 June 2026 (UTC)[reply]
I think I understand what you say, but I feel like that's besides the point I was trying to make. Is the first check ("First check for the AI policy") a scope check, or at least a part of one, deciding in-scope vs out-of-scope? If so, how do you explain that it is also supposed to be one of the "reasons beyond their scope" according to the proposal at the same time?
Should we make the new rule a separate document from the project scope policy and reference it from the scope policy, as a matter "beyond their scope"? That seems more straightforward to me. Is there any reason why AI generation has to be a scope issue? whym (talk) 03:04, 26 June 2026 (UTC)[reply]
Whym, it's an additional requirement. For example, take "Must be an allowable free file format". There are typically technical restrictions to prevent files in unfree formats from being uploaded at all. But let's say you manage to do it anyway. Even if your file has educational value and is in use and is freely licensed, we still delete it if the file uses an unfree format. And not being freely licensed is an existing "reasons beyond their scope" bullet point, even though the requirement to be freely licensed is also covered by COM:Project scope#Must be freely licensed or public domain. I do understand this might be confusing for some, so this can be discussed on Commons talk:Project scope#"reasons beyond their scope". - Alexis Jazz ping plz 00:09, 28 June 2026 (UTC)[reply]

Use of authentic/inauthentic

[edit]

The proposal unnecessarily implies that non-AI media are "authentic" and AI media are "inauthentic". This is problematic for several reasons:

  • Authentic/inauthentic are not defined
  • The proposal unnecessarily conflates authentic/non-AI and inauthentic/AI; whereas some AI-generated/modified media can arguably be authentic (e.g., an svg diagram); and some non-AI-generated media can arguably be non-authentic (e.g., a hand-drawn picture of a dinosaur)
  • The wording implicitly frames AI-generated content as deficient or deceptive, embedding a value judgement into what should ideally be a neutral descriptive policy
  • Policy should focus on observable characteristics rather than debatable philosophical concepts

Instead, use language that is uncontroversial - for example, for the first sentence:

"The following kinds of AI generated files are not allowed on Commons unless it is indicated in-image and discernible in a 250 pixels wide thumbnail that the work is not authentic."

change to:

The following kinds of AI-generated files are not allowed on Commons unless it is indicated in-image and discernible in a 250 pixels wide thumbnail that the work is AI-generated."

Similar wording changes are suggested for each of the subsections.

-- Jtneill - Talk 10:15, 24 June 2026 (UTC)[reply]

How to check a 250px thumbnail

[edit]

What practical methods can a contributor use to check whether a watermark is clearly visible at 250 pixels? Could the proposal point to additional guidance?

-- Jtneill - Talk 10:25, 24 June 2026 (UTC)[reply]

Examples

[edit]

Could links to examples be provided for each proposed section? Contributors would likely benefit from seeing concrete examples that illustrate how the requirements are intended to be applied in practice. -- Jtneill - Talk 10:32, 24 June 2026 (UTC)[reply]

Machine-readable identifcation of images having this watermark

[edit]

Please include some sort of machine-readable indication that a watermark has been added to the image, so wikis with "no watermarks" policies (e.g. w:en:WP:WATERMARK) can better audit when images from Commons will need to be replaced with some alternative. For example, all such images might include a specific template or be placed in a particular category. Anomie (talk) 12:53, 24 June 2026 (UTC)[reply]

I would suggest to add a parameter to the {{PD-algorithm}}. Something like a "comply=yes" if the file complies with this policy. GPSLeo (talk) 15:29, 24 June 2026 (UTC)[reply]
Would "comply=yes" be specific to watermarked images, or would people also use it for images where the "reasonable person" tests don't mandate a watermark? If the latter, that would not be sufficient for the auditing mentioned here. Anomie (talk) 01:08, 25 June 2026 (UTC)[reply]
enwiki policy prohibits the use of most AI-generated images, outside of noteworthy AI-generated images and samples of AI output. Someone should start a discussion on enwiki about making an exception to the "no watermarks" policy for AI disclosure watermarks; I think they'd be open to it. Omphalographer (talk) 23:01, 25 June 2026 (UTC)[reply]

AI upscaling has been the default for years

[edit]

Overall it's a good initiative. I wouldn't want to be in a situation when I open a category on Commons and I can't easily say which images are have been substantially altered. So everything here is my feedback to make the policy more effective.

Upscaling has been applied as the default for years [1], way before generative AI became a thing. Commons has thousands and thousands of such photos taken by phone cameras in the default mode. We probably don't want to remove or tag them so probably we need to define upscaling in a more precise way. Alaexis (talk) 19:44, 24 June 2026 (UTC)[reply]

(I probably should go back and edit my vote on that section) That is a good point, wording probably should more precise. Abzeronow (talk) 02:22, 25 June 2026 (UTC)[reply]
Alaexis, oh my God, I feel vindicated.. For years, when comparing the low-light photos from GSMarena I thought the iPhone started hallucinating, starting with the 11 Pro, particularly visible on the London bus map. It kept getting worse every generation and only seemed to affect the low-light photos. Starting with the 15 Pro it got better again. This is actually why I bought an iPhone XS not that long ago and never got anything from the 11-14 generations, but I'm considering the 15.
Now you're telling me I wasn't just seeing things, it's real. The hallucinations aren't on the level of your average chatbot, I'm a pixel peeper, but they are there. Whether this should be categorized as "AI" is a good question. I wouldn't really recommend a completely blanket exemption for this. For the iPhones, yes, exempting them would be reasonable if only Deep Fusion was used (which can't be turned off, other than by getting sufficient light). And this is probably true for most if not all smartphones today. It's unclear what the future will bring though, so a blanket exemption would be unwise in my opinion.
It's unclear if this could/should be categorized as AI, but if this RfC passes I can make a proposal on Commons:VPP to ensure these cases are exempt, which should be a slam dunk. I don't think a complete RfC would be needed for that. - Alexis Jazz ping plz 00:51, 26 June 2026 (UTC)[reply]
You may want to read about the Samsung Galaxy S21 moon photo controversy as well. Apocheir (talk) 02:35, 26 June 2026 (UTC)[reply]
I agree. I raised my concern above. If AI upscaling or minor editing is banned or substantially burdened, I'll just upload my crappy photos and let someone else fix the photos for me. OhanaUnitedTalk page 19:38, 28 June 2026 (UTC)[reply]
OhanaUnited, there's no reason to use AI upscaling, other than to demonstrate AI upscaling itself. For minor AI editing, you'd have to upload the original as well. It's the only way it can be verified that the alterations the AI made were truly minor. If that's too much effort and as a result you will only upload the originals and "let someone else fix the photos", that's good! Like most here, you're just a volunteer. You have no obligation to do anything. I fix photos from other people all the time. If you provide the original, we can always fix anything that needs fixing later.
You voice specific concerns over File:RWDI Guelph headquarter office.jpg. You could overwrite that file with the original, then revert to your preferred AI-version. (Canada law is okay with people who are included incidentally) We have no requirement to use {{Personality rights warning}} and generally don't actively add it for a person who was incidentally included in a photo.
You say you don't know how to Photoshop. That's okay. For those cases where you really can't publish the original at all, would you be willing to learn how to draw a rough selection and either blur it or fill it with one color, which can also be done with free software? Generally it's as simple as finding a rectangular selection tool and making a rectangle that covers the unwanted. Press "delete". Or, for a nicer looking result, find "Gaussian blur" in a filter menu. Instructions for GIMP.
Someone else could perform a nice content-aware fill or clonestamp the unwanted elements out at a later time, if needed. - Alexis Jazz ping plz 20:45, 28 June 2026 (UTC)[reply]

A faulty premise

[edit]

"Large numbers of images that do not reflect reality are being uploaded. When an inferior AI image is inserted into an article it disincentivizes actual photographers and graphic artists from creating images because they are more inclined to create an image for an article that has none. AI-altered images make the detection of violations of copyright more difficult." The proposal is based on this assumption, which takes several inaccuracies for granted

  • That photographers and graphic artists actually have that overly specific motivation.
  • That the work of photographers and graphic artists is superior to that of AI, just because
    • That the work of photographers and graphic artists "reflect reality", merely because they have a human author.
  • That the lack of an image will motivate someone to make an image to fill that void.
  • That photographers and graphic artists are always readily available to illustrate an article in need... or worse, that anyone can be a photographer or graphic artist to do so.
  • That the number of AI images being uploaded is indeed large.
  • That fanart is a new thing introduced by AI and that Commons is easily fooled by it.

This proposal is just an anti-AI witch-hunt, frequently seen on several internet pages. It is not based on reasonable arguments, just fear monguering. Cambalachero (talk) 19:55, 24 June 2026 (UTC)[reply]

While I agree that no evidence was provided for load-bearing claims, I think that the problem is real. GenAI makes it possible to easily create unlimited number of artifacts, so absent some kind of filtering Commons can be swamped with AI-generated media. Alaexis (talk) 20:14, 24 June 2026 (UTC)[reply]
Quantifying how many files might be affected by new requirements would be a good step. It's tricky since AI 'detectors' are fallible, but even a ballpark estimate could help determine the scale of this RfC. Anecdotally, it's not hard to find examples which might fall on one side or the other of a given line in the sand. For instance, is file:POV of an ant walking inside an ant nest tunnel (Sora AI-video).webm a helpful visualisation which as the file description notes would be difficult or even impossible to capture in the wild, or is it just pseudozoological fiction (note the four-legged ant)? — Arlo James Barnes 20:53, 24 June 2026 (UTC)[reply]
I personally review at least a few dozen unambiguously AI-generated images uploaded each day, and I suspect there are a lot more that I don't see. What you see in my DRs isn't even the whole picture; a lot of AI-generated uploads are also clearly personal content (e.g. self-uploaded AI-generated profile photos of aspiring musicians or entrepreneurs) and are speedily deleted under CSD F10. Omphalographer (talk) 20:37, 25 June 2026 (UTC)[reply]

' Agree' - the premise is unfounded, ideological nonsense, or if you prefer "opinionCRUFT". Lx 121 (talk) 10:46, 26 June 2026 (UTC)[reply]

test case

[edit]

In Hebrew Wikipedia we have discussion about her.

  1. File:Yaʻaḳov Yitsḥaḳ Yuṭes (Published in 1884).jpg - this is the original file on Commons
  2. HE:קובץ:יעקב יצחק יוטס.png - this file is AI modified
  3. HE:קובץ:יעקב יצחק יוטס 1.png - this file cropped is AI modified
Comments in the he.wiki: The AI ​​images not only cleaned up the original image, but also recreated structural features of the face that were unclear. For example, the nose was rebuilt. Since the original image doesn't give much information about the structure of his nose, AI tools can't know what the nose really looks like, instead they use statistical inference, which is based on patterns learned from training on databases with many other images. (@Politheory)
Beyond Politeori's comment that there are indeed changes to the facial structure that are not faithful to the original (which can also be argued for Photoshop retouching), the Achilles heel of AI in images is working at the pixel level without a vector understanding of letters. Because the model draws the visual form of the text instead of encoding it as a logical symbol, it distorts the caption and invents random fonts that look visually correct and sometimes even linguistically meaningless.
Indeed, that's what I meant - looking at the AI ​​image with the caption raises the question of what else the AI ​​accomplished in making the image. (@Szazal)
Changed the eyes, added forehead features, nose as Politheory noted.
In some cases, this is no different from the retouches' who do it in Photoshop, who also stretch, smooth, and change facial features and eyes, etc.(user:Geagea)
The main problem is with the change to color, where it is impossible to know at all if it is even close (perhaps with in-depth research and prior knowledge of the person, it is possible to approximate it by simply giving appropriate instructions). (User:Szazal)

opinions — Preceding unsigned comment added by Geagea (talk • contribs) 10:00, 26 June 2026 (UTC)[reply]

Notifications: @Politheory, @Szazal. -- Geagea (talk) 12:40, 26 June 2026 (UTC)[reply]
Except for the eyes not looking quite the same direction, (2) is a plausible interpretation of (1). I agree with the remark above that the nose is a bit conjectural. (3) simply is not a plausible interpretation of (1), it's a highly conjectural portrait informed by (1).
I could imagine some educational value to (2), if properly captioned to be clear that it is imperfect and conjectural. (3) seems to me like a waste of pixels. - Jmabel ! talk 00:40, 27 June 2026 (UTC)[reply]

Obvious Problem

[edit]

-WHEN the original open-source licensed file DOES NOT have a watermark, & we ADD a watermark, how do we represent the ORIGINAL FILE? or is this an elabourate rational for quick & easy DR's? (my apologies if this has been addressed above somewhere, but i could not find any extensive consideration of it) Lx 121 (talk) 15:10, 26 June 2026 (UTC)[reply]

@Lx 121: You can always upload an original version and then promptly overwrite it. In other circumstances, we do this a lot. See File:Roman portrait head of a man of rank - Staten Island Museum - 03.jpg for a recent example in one of my uploads. - Jmabel ! talk 00:44, 27 June 2026 (UTC)[reply]

PROBLEMATIC WORDING

[edit]

a lot of the wording in the initial proposal is problematic, & some of it at least seems ideologically driven, like these:

  • 'the work is not authentic' - howso/define "authentic"/"inauthentic" in this context?
  • 'Large numbers of images that do not reflect reality are being uploaded' - COMMONS is full of "images that do no reflect reality"; fictional, hypothetical, graphic arts, etc. - this is within common's scope, ergo, NOT in itself a valid rationale, w/o qualification.
  • 'When an inferior AI image is inserted into an article it disincentivizes actual photographers and graphic artists from creating images because they are more inclined to create an image for an article that has none.' - as discussed above, the claimed premise id DEEPLY flawed, but ALSO, re: "inferior" - not all ai images are "inferior", so how is the user qualifying & determining this? because it reads like a sweeping (& inaccurate) generalisation (which takes us back to the "ideological"/manifesto aspects of this proposal).
  • 'AI-altered images make the detection of violations of copyright more difficult' - Rubbish; it is literally NO DIFFERENT than detecting & determining copyvio in ANY OTHER image. the only way you get anywhere CLOSE to such a claim is if you buy in to the highly dubious ip claims around training material, which are almost entirely UNTESTED in court, & mostly used to extort settlement deals. IF it ever makes it through the trial process in us courts, & the supreme court doesn't garbage it, THEN it becomes a "valid" rationale. UNTIL that hypothetical "then", wikimedia has pretty strong history of defending the rights to free use.

Lx 121 (talk) 10:46, 26 June 2026 (UTC)[reply]


HERE'S what is wrong with this whole approach:

[edit]

1. There is NO DIFFERENCE, fundamentally, between the image creation & editing techniques that AI uses, & other EXISTING techniques that are Accepted on Commons. - AI just makes them easier.

2. AI is here - it exists, it is VERY good, & it is getting better FAST, iteration by iteration. in ~10 years, AI will probably be BETTER than humans are at a WIDE range of tasks. in ~3-5 years most of the problems/complaints with AI-generated materials (that are under discussion here on commons) will be largely fixed. - AI is evolving faster than commons can make policy for it.

3. What other types of material will we start "requiring" watermarks for, ONCE the precedent is established? - commons "doesn't like" watermarking for GOOD REASONS. forcing this on AI files is overkill. including suitable entries about it in the STANDARD file info is the appropriate, & established practice on commons. AND it should include detailed info - what was done: created/edited, - which AI, etc. JUST LIKE we already do for other image sources, image edits, derived works, etc.

The tl,dr - AI image work is not a "special" new category of thing. It is just the SAME TOOLS that we already have & use, only faster & easier to use.

Supplementary:

- the "ludditism" on here is particularly ironic given the VAST not just "array", but LAYERS of automatic tools we use on her to make mediawiki easier to deal with.

- the "tools" used to identify ai created/edited materials on here are crude at best. mostly they can't even tell the difference between a minor edit & a "from scratch" image generation.

the fuss over ai created/edited material is something like a "moral panic" (the last one i strongly noticed was ~10 years ago, & had to do with use of certain types of wmc images to vandalise wikipedias; & the issue is pretty much resolved now & looks more than a little overblown in retrospect)

Lx 121 (talk) 11:29, 26 June 2026 (UTC)[reply]

There IS a fundamental difference between AI image generators and other tools: the scale of production. You said it yourself. But I think the main reason this RFC has been opened, is not because of a supposed moral panic, it is just that most people simply don’t like AI-generated stuff for a lot of reasons and want to be able to spot such stuff more easily to avoid using it entirely:
  • AI-generation tools are giant plagiarism machines which regularly infringe copyright;
  • They are very bad for the environment;
  • They generate stuff that is generally considered ugly;
  • They make it easier to create propaganda/misinformation;
  • etc.
These points concern both contributors and readers alike, and I seem to remember that transparency towards our readers is a core ideology of our wikis, isn’t it? :) Danÿa (talk) 20:35, 26 June 2026 (UTC)[reply]
  • AS is NOT getting better over time (maybe bigger and more resource-hungry, but that's a different story)
  • AS is bad for the legal system (no photo or video of evidence has any value anymore)
  • AS is bad for education, schools are desperate like never before
  • AS is VERY BAD for the environment
  • AS is bad for wikimedia
  • AS is bad for society (ie mass fake news and mass harassment as the most popular applications)
  • AS is bad for your brain (like the car is bad for your legs)
Still love AS? No problem: here is your spot, NOT at wikimedia. Taylor 49 (talk) 00:44, 27 June 2026 (UTC)[reply]
  • Actually, there are several a deep differences between AI editing techniques and those in a program such as GIMP. When I use GIMP to alter an image, I have virtually complete control of the changes made; about the only thing beyond my control is where interpolation comes into play, and even then GIMP is consistent and the results are entirely reproducible. When someone uses AI to alter a comparable image, and requests a comparable alteration, AI may insert an open-ended number of artifacts of many different types, may affect portions or aspects of the image it was told not to alter, etc., and its response to a given prompt is not reproducible. - Jmabel ! talk 00:52, 27 June 2026 (UTC)[reply]

We probably don't need AI-upscaled image on Commons

[edit]
Moved from #AI upscaling: 1F616EMO (talk) 05:01, 27 June 2026 (UTC)[reply]

I don't really see any point to allow the inclusion of AI-upscaled image on Commons, as they don't add any new educational values while contribute chaos to the Commons due to their inauthenticity. If some individual wants to upscale, they can do that themselves locally; if a wiki wants to upscale, for example, to demonstrate how upscaling work, they can upload locally. But if consensus decides that we should allow AI-upscaled images nevertheless, then I support this provision. 1F616EMO (talk) 17:34, 23 June 2026 (UTC)[reply]

An obvious exception would be when the upscaled version is the only version found. We should do our best to find the original one, yet in the meantime, we may upload a watermarked version of the upscaled image. 1F616EMO (talk) 17:36, 23 June 2026 (UTC)[reply]
Why I am opposed to AI upscaling but not other AI-generated works: As long as editors refrain from dumping all outputs from their AI models (like how we won't dump our whole camera roll onto Commons), AI-generated works are reasonably educational considering the human knowledge involved and efforts paid in prompting and refining, as well as the final selection of images among all the outputs. Upscaling, however, does not add any new values in comparison to the AI-free original version. 1F616EMO (talk) 17:47, 23 June 2026 (UTC)[reply]
WRT "upload locally" - a significant number of wikis, including some large ones like eswiki, do not support local uploads. See Commons:Turning off local uploads for details. Omphalographer (talk) 21:48, 23 June 2026 (UTC)[reply]
It is out of our scope to host fair use images even if some wikis do not have local uploads; likewise, it is out of our scope to host images with little or no educational values even if some wikis do not have local uploads. By disabling local uploads, wikis accept that the catalogue of images they can choose from is fully governed by the Commons, and therefore subject to our scope, policies, and guidelines. If they want to use images out of our scope, they should enable local uploads, possibly lock it behind administrator rights, and upload those images locally. 1F616EMO (talk) 02:42, 24 June 2026 (UTC)[reply]
Commons already has thousands of ai-upscaled images because AI upscaling has been the default in most mobile phone cameras for many years already. Alaexis (talk) 08:30, 27 June 2026 (UTC)[reply]

AI-assisted images

[edit]

So there has been a lot of discussion about AI-generated images (not just in this thread, but on Commons as a whole), however there has been little talk about AI-assisted images. They have no categories or templates and I haven't seen any notable examples of AI-assisted images (mostly due to copyright, but freely licensed assisted images could also be done). How should these be handled?

If one generates an AI image, traces it in SVG to recreate it in vectorized form without artifacts and fix blatant errors, then uploads it to Commons, would this be allowed? I'm asking this example specifically because I'm planning to do exactly that. Now I will tag the image as AI-assisted upon upload, provide info such as prompt and seed, and may even provide the unedited AI image as reference and proof, but do you think it should it be required, to provide this info?

But what if someone just airbrushes a few things off a clearly AI image and calls it a day? These minor changes are incomparable to full vectorization described earlier, but would it still be allowed? Where should the line be? Dabmasterars [EN/RU] (talk/uploads) 21:06, 27 June 2026 (UTC)[reply]

RFC is flawed as it forces visible watermarking on files

[edit]

The accepted Commons practice is that visible watermarks are unacceptable or discouraged, Commons:Watermarks. There's no real necessity to damage AI enhanced or AI created images of educational value that are acceptable under future AI related Commons policies. If there must be an AI mark it could be added as an additional chyron type bar in sympathetic colour rather than damaging the original image, diagram or video. The RFC needs to be re-written so the consensus is based on what is an acceptable "watermark" or other type of non-visible flag or tag which could automatically display the file on Commons with a label visible to potential reusers. It could even be added as a license condition if it's to be a policy to avoid misuse downstream by reusers.

The response that versions that are not watermarked might be in the file history is not a solution for 95% of casual reusers. Most of reusers have no idea that the file history has downloadable items.

If this RFC passes, consideration should be given to a secondary RFC to clarify what is to be enforced as "watermarking" and to ensure valid alternatives that do not mean Commons is a host to avoidably watermarked educational material. --Fæ (talk) 08:15, 29 June 2026 (UTC)[reply]

I agree. This proposal is ideal for all those who wish to purge as much AI-generated or AI-altered content from Commons as possible. Even though the proposal is not entirely wrong, it is perceived by me as an attempt to find very simple answers to some quite difficult questions. I sincerely hope that the proposal does not pass without substantial improvements. Watermarking makes images essentially useless and it is certainly not acceptable in Quality Images (QIs). With the current proposal, QIs taken with a phone would be essentially impossible, not just rare as nowadays, due to AI upscaling by the phone that cannot be influenced by the photographer. Most people do not own cameras any more, but they take photos with their mobile phones. Who would actually upload such content to Commons if they needed to add a visible watermark to each and every photo? In addition, and if I am not entirely mistaken, there may be thousands of QIs of birds and butterflies that were altered by AI-sharpening and AI-denoising without disclosure. Of course, the respective authors could upload all the original versions and link them from the image pages of the altered versions, but I strongly doubt that this is going to happen.
I personally marked all of my AI-edited images with a template, in my case {{Created with Topaz DeNoise AI}}, and I uploaded the "non-AI" versions for most of these images. However, there are a few very early examples, for which I failed to do this. Finding all of these might be the trickiest challenge that I would have personally with this proposal. I would probably just overwrite the respective images by the original photos from the camera and then revert this action.
There might be yet another issue for me. One of my cameras is known to produce images that may be a bit blurry, but are known for rather strong internal denoising and sharpening by default. I reduced the sharpening as much as possible. Loss of detail by internal denoising is still possible, though. I bought the camera in 2015. The term "AI" was not as omnipresent then as it is today. However, if someone claimed that the camera does something "AI-based" to the photos, all the photos from this camera might become unacceptable without watermarks. Simple answers to difficult problems usually get a lot of approval nowadays, but they are almost always wrong. --Robert Flogaus-Faust (talk) 14:52, 29 June 2026 (UTC)[reply]

I don't get it

[edit]

If someone wants an AI illustration for a certain topic, they can simply prompt an AI tool to generate one. Why then do such files need to be stored on Commons? Even if many of them may be (arguably) educationally useful, it will always be possible to generate e.g. an image that is better in at least some way, depending on what it's needed for. Please note that I'm not talking about AI-edited images or similar situations. Sinigh (talk) 09:16, 29 June 2026 (UTC)[reply]

If is actually in scope and even in use, then it should be on Commons IMO. Many AI-generated images are out of scope even now without this policy update. By the way, your reasoning could be applied to many photographs as well. There could always be a better one in many cases. --Robert Flogaus-Faust (talk) 13:02, 29 June 2026 (UTC)[reply]
My point is that AI images can be generated freely at anyone's convenience and thus do not need to be supplied by a media database. Sinigh (talk) 14:07, 29 June 2026 (UTC)[reply]
They have to be on Commons to be used on Wikipedia or other WMF projects (unless the project allows local uploads). If it makes sense to have that illustration, then it makes sense to have it available on Commons, instead of editors needing to recreate it (which may require substantial work) or find it on an odd Wikipedia – in addition to enabling local uploads. The question is whether it makes sense to use such an AI-generated illustration in the first place, or in what situation it makes sense. –LPfi (talk) 14:09, 29 June 2026 (UTC)[reply]
Retrieved from "https://commons.wikimedia.org/w/index.php?title=Commons:Requests_for_comment/Policy_update_for_AI_content&oldid=1240894372"