Jump to content

Wikipedia talk:Image use policy

Page contents not supported in other languages.
From Wikipedia, the free encyclopedia

AI-generated images of people

[edit]

I came across this 2023 archived discussion about AI-generated images, but it doesn't appear to have reached a clear-cut conclusion, and it also appears to have been mainly focusing on the copyright status of such works. There are encyclopedic concerns as well, particularly when it comes to AI images of people. I occasionally come across what seem to be AI generated images of people while looking at Special:NewFiles: some recent examples are File:Bafaki Tangal.jpg, File:P. M. S. A. Pukkoya Tangal.jpg, File:Km sahib.jpeg, File:Sayyid ummar bafaqi.jpeg and File:K. Uppi Saheb 1.jpg. To be honest, I don't know for sure whether these are AI images, but they don't seem like paintings, drawings or photographs. From a copyright standpoint, these could be problems per c:COM:BASEDONPHOTO if they were created based on a old photo or something; however, even if they're 100% original, they might be too original per WP:IMAGEOR. The question I have is whether encyclopedically such images are OK to use even if their copyright is not a problem. FWIW, there is a little on AI images in WP:AI#Images and Commons and much more in c:COM:AI, but these too seem more focused on copyright related issues than encyclopedic use; of course, copyright is what Commons is more concerned with, which is why it might be a good idea for encyclopedic concerns covered a bit more locally here on Wikipedia. The two images I recently came across are of deceased persons and its possible non-free images could be used per WP:NFCCP; if, however, that could be affected if freely licensed AI-generated images are considered to be a reasonable free alternative to non-free images. Furthermore, freely licensed AI-generated images could possibly be argued to even be OK to use in BLPs, but that might cause issues with WP:BLPIMAGE. Pinging Masem and SMcCandlish since they participated in the archived discussion mentioned above, but feedback from others would be appreciated too. -- Marchjuly (talk) 01:24, 8 October 2024 (UTC)[reply]

I think it's trivially obvious that no AI-generated image has any encyclopedic value whatsoever. This should be enshrined in policy. Encyclopedic value is a function of human discernment in collation, preparation, and representation. AI is fundamentally incapable of such discernment; as AI is functionally a black box, humans purporting to mediate its output are also incapable of such discernment. Remsense ‥  01:25, 8 October 2024 (UTC)[reply]
I know the scope of this post is intentionally narrower, pertaining to images of people. However, I do not see a distinction worth making here. Remsense ‥  01:43, 8 October 2024 (UTC)[reply]
I could see some value in using AI generated images in articles about AI generated images or maybe in articles about art, but I'm not so sure there's much value in using them in biographies, except perhaps as an example of someone's art or perhaps in cases where the image is controversial and the subject of critical commentary in reliable sources. Given that so many biographies (not only BLPs) seem to be without at least an image for primary identification purposes, the temptation to create one using AI could be too much for some to resist. Moreover, some non-free images might be of poor quality (File:K. Uppi Saheb 1.jpg was actually overwriting File:K. Uppi Saheb.jpg and needed to be split, and File:Bafaki Tangal.jpg is another overwritten file in need of a split.) that it's tempting to replace them with "better" looking AI images. The questions is whether such a thing is good from an encyclopedic standpoint. If the consensus is that it's not, then I agree such a thing should be clearly stated in relevant policy pages. -- Marchjuly (talk) 01:47, 8 October 2024 (UTC)[reply]
If we have free non-generated images we should use those. If we do not, then works derived from non-free images, whether created by AI or by a human artist, are likely problematic with respect to the copyright of the images they were derived from. —David Eppstein (talk) 01:54, 8 October 2024 (UTC)[reply]
Yes, I wasn't sure how to articulate this while coming off with adequate clarity as to my position, but it's clearly reasonable to use AI-generated images as primary illustrations of the generation itself. Remsense ‥  01:58, 8 October 2024 (UTC)[reply]
Zero allowance for AI generated images that are meant to depict people, living or dead. Within the context where AI images would be allowed and where it is needed to show a human or more with the image, it should be clearly generic human figures that AI is known to generate, and if the image edges on recognizability, an alternative image should be saught. — Masem (t) 02:16, 8 October 2024 (UTC)[reply]
@Masem: Does it already say that somewhere in IUP or some other policy page? Is there fairly accurate way of determining whether an image is AI-generated? I'm a total newbie when it comes to them and only notice when the image seems unnatural for some reason. I'd imagine that some are quite skilled at creating such images so that they can be really hard to detect, unlike the ones I mentioned above (which really seemed odd to me). -- Marchjuly (talk) 02:25, 8 October 2024 (UTC)[reply]
Sorry, I should have said this is my opinion , if we don't have advice anywhere on this. — Masem (t) 02:30, 8 October 2024 (UTC)[reply]
I think this is an argument for expecting the sourcing and processing to be clearly stated on the file page for media used in articles. I've seen people point to the frequent absence of this necessary documentation on Commons as a pragmatic argument that we can't truly expect WP:V to always apply. I personally won't apologize if I'm doing GAN or peer review for an article where I have to insist on removing media with incomplete or unclear documentation. Remsense ‥  02:32, 8 October 2024 (UTC)[reply]
Yeah, I think if we are allowing AI images (and in such cases, starting with how Commons does it), it should be where the uploader has otherwise been in control of the image generation route so we know what the prompts were, what AI engine was used, etc.
But again, that's for sufficiently generic images. When it comes to any AI image that tries to produce images of known persons, that should be an area we avoid with a ten-foot pole due to the potential issues with accuracy, representation, etc. Masem (t) 03:04, 8 October 2024 (UTC)[reply]
I find myself in concurrence with all of the concerns raised above about use of AI-generated (or significantly AI-altered) images of people, outside the context of encyclopedic coverage of what AI imagery is and what controversies surround it. If we're using AI fakery to represent biographcial subjects, then we are making a mistake. As for identifying them, there are sites now that analyze images and can identify AI-generated ones with a 95%+ accuracy rate, so I'm told, though this is not something I have looked at closely.  — SMcCandlish ¢ 😼  05:53, 8 October 2024 (UTC)[reply]
Anyone uploading an image like that needs one warning and then an indefinite block. At the moment, the couple of samples I've seen are hideous (example from above). However, even if that problem were overcome, the idea that faked photos could be used because someone thinks they are ok shows a clear WP:CIR problem. Johnuniq (talk) 06:13, 8 October 2024 (UTC)[reply]
That's what makes me confident that expecting clear documentation for images used in articles is adequate to address this problem on the article front at least: or, at least addressing it as well as we address copyvio in prose. Editors who do this are almost always inexperienced or incompetent, and scandals where a regular contributor in good standing is found to be fabricating this documentation are likely to be exceedingly rare.. Remsense ‥  06:18, 8 October 2024 (UTC)[reply]
FWIW, I didn't start this discussion in the hope that someone would drop the hammer on one particular uploader, but rather to see whether this kind of thing had been previously discussed before. I'm assuming the uploader of the files I used as reference is just unfamiliar with such things and was acting in good faith. They were overwriting existing files with some of their uploads which is also a problem, but again I think this was just a new user not familiar with such stuff. I used their uploads for reference because they're ones I recently came across in Special:NewFiles, but I've seen such images before and always wondered about them. Such images looked odd to me but I wasn't sure how to tell whether they were AI generated. I imagine these types of images are going to become more and more common on Wikipedia since they seem to have become more and more common out in the real world. So, if there's no existing policy like IUP or BLP that specifically deals with them, then perhaps it would be a good idea to discuss one. If a consensus is established and there's a way to technically detect them, then existing images could be taken care of and perhaps steps could be implemented to prevent/discourage future uploads. -- Marchjuly (talk) 07:38, 8 October 2024 (UTC)[reply]
I think BLP at least in spirit covers it. Fundamentally, an AI-generated image misrepresents a real, living person, showing them in a way they never actually looked like. If it's not acceptable to make misrepresentations about living people in text, it certainly should not be acceptable to do it in images either. Seraphimblade Talk to me 11:09, 8 October 2024 (UTC)[reply]

Since this discussion has yet to be archived, I figure I'll ask about File:Stephen A. Love.jpg and File:SAL LINKEDIN pp.jpg. Do these appear to be AI-generated or AI-enhanced? They could've been uploaded by Stephen A. Love himself. Would that make a difference if true? -- Marchjuly (talk) 07:45, 5 February 2025 (UTC)[reply]

The yellow one claims to have first been published in 2014, i.e., before AI photo tools were generally available. WhatamIdoing (talk) 08:00, 5 February 2025 (UTC)[reply]
The second one has metadata and is genuinely on LinkedIn. CMD (talk) 09:25, 5 February 2025 (UTC)[reply]
Could AI be used to enhance an older image? -- Marchjuly (talk) 10:10, 5 February 2025 (UTC)[reply]
Could be, but there are also other ways to enhance an image. Either way, it seems this is an image used by that person to represent themselves, which is not the same as someone else creating an image of them. (Whether the image permissions are appropriate is a different matter.) CMD (talk) 15:10, 5 February 2025 (UTC)[reply]
What has AI got to do with it? The first appears to be from an album cover. In the second perhaps the guy has whitened his teeth with an image editor. Why does it matter if this was done with an AI tool or a "dumb" image editor. This appears to be AI paranoia. -- Colin°Talk 10:40, 5 February 2025 (UTC)[reply]
Marchjuly, it looks like you have not seen Wikipedia:Village pump (policy)#Guideline against use of AI images in BLPs and medical articles? yet. Warning: it's already over 350 comments. WhatamIdoing (talk) 01:19, 7 February 2025 (UTC)[reply]

Deletions

[edit]

Images were deleted here w the edit summary "decorative and unnecessary." https://en.wikipedia.org/w/index.php?title=74th_Street_(Manhattan)&diff=prev&oldid=1265766246

Is it permissible to restore some or all of them? I can't find support for the deletion in our policy or practice.

If the answer is no, is it then permissible to delete the similar images at List of people from Chicago, List of people from Houston, List of people from Kansas City, Missouri, List of people from Ottawa, List of people from Berlin, List of Catholic priests, List of African-American activists, and the like?

Interested in thoughts here. Thank you. 184.153.21.19 (talk) 23:11, 8 January 2025 (UTC)[reply]

Hi, I'm Jane Fonda on a beach someplace. Should I be in a Wikipedia article about a street in New York?
Photos of people were removed from the "notable people" section at 74th Street (Manhattan). Wikipedia is not an image gallery. If readers want to see this photo of Jane Fonda--which is decorative and has no relevance to that particular article--they can read her article. Magnolia677 (talk) 23:28, 8 January 2025 (UTC)[reply]
I don't imagine that that particular image is the issue, but if it is one can no doubt find a different one. --184.153.21.19 (talk) 23:35, 8 January 2025 (UTC)[reply]
You're missing the point. Magnolia677 (talk) 23:57, 8 January 2025 (UTC)[reply]
The relevant guidance is MOS:IMAGEREL. 𝕁𝕄𝔽 (talk) 23:01, 20 April 2025 (UTC)[reply]

Displayed image size

[edit]

In view of m:Tech/News/2025/16 (currently posted at Wikipedia:Village pump (technical)#Tech News: 2025-16 and some user talk pages), I think that we should update WP:THUMBSIZE to explicitly name these preferred standard sizes, whilst keeping the recommendation to use |upright=, in order to discourage unnecessary sizing. In the last couple of years I have noticed increased use of strange image sizes that bear no relation to the original upload size; often, they are odd numbers. See e.g. this edit where I amended the dimensions of five images, one of which specified |413x413px. --Redrose64 🌹 (talk) 08:05, 15 April 2025 (UTC)[reply]

Those HxW dimensions are a result of people resizing the image in the visual editor. Sjoerd de Bruin (talk) 12:18, 15 April 2025 (UTC)[reply]
Oppose. The scaling factor for upright should be updated to use 250 instead of 220 and a mention of 250 being the default size, not 220. Thumbsize already discourages using specific sizes. Snævar (talk) 13:52, 15 April 2025 (UTC)[reply]
  • That guidance should be updated, since static sizing is deprecated. Nikkimaria (talk) 13:43, 20 April 2025 (UTC)[reply]
    200px also appears to be from a much earlier era of screen size - it's a good width on a 1024p screen, but too small on anything bigger. (Well over half of desktop screens worldwide are 1366p or bigger). Wikipedia UI should reflect what users will actually see - pretending that no one has replaced their 1024p screens since 2008 doesn't make sense. If the UI had kept pace with screen size, we would have thumb sizes larger than we currently do. Pi.1415926535 (talk) 05:20, 21 April 2025 (UTC)[reply]

Lead image size

[edit]

I just reverted a change saying that we should reduce the upright= parameter on lead images from max 1.35 to max 1.2, so that with the default image size (now 250px rather than 220px) the max pixel dimension of the lead image stays at most 300px. This is going to affect many articles that in good faith used the previous value of upright=, and also affect many readers who (like me) bump my default image size from the system-wide default. I think it needs a discussion. Does the lead image size setting need to be 300px for some technical reason (in which case we should encourage setting it in lead images to 300px rather than by upright=), should it remain proportional (in which case I think the previous advice of max upright=1.35 serves to keep it from being too oversized), or is there some good reason for continuing to encourage proportionality using upright for lead images but reducing the proportion to 1.2? There was some vague wave at infoboxes in the edit summaries but as far as I can see that is irrelevant because when there is an infobox the lead image usually goes in the infobox and infobox lead images are usually not proportioned upward with upright=1.35, so the articles with upright=1.35 leads and the articles with infoboxes are almost entirely separate from each other. —David Eppstein (talk) 18:32, 19 April 2025 (UTC)[reply]

I believe that this is a continuation of the discussion above. This page currently states that the default thumbnail size is 220px, which is incorrect. At a minimum, we should not have incorrect statements on policy pages. – Jonesey95 (talk) 19:01, 19 April 2025 (UTC)[reply]
However standard Wikipedia "infobox" image size is supposed to be 200px, lest they crowd out the lead text. That's a mess now. Fyunck(click) (talk) 19:47, 19 April 2025 (UTC)[reply]
We should change the statement of the default image size to be correct. But the discussion of the max image size for non-infobox lead images (status quo ante: upright=1.35), and the image size for infobox images (status quo ante 200px?), are two different things from the default image size and if we want to change them from the status quo we should discuss them. —David Eppstein (talk) 20:16, 19 April 2025 (UTC)[reply]
@David Eppstein: I don't want to change them from the status quo. That was fine. All I'm saying is that with the change to 250px for thumbs it also broke all infoboxes... or at least the sports infoboxes. It overrode the standard infobox 200px. Countless 10s of thousands of infoboxes that were just fine are now broken as gigantic. I wonder if we could hard-code the infobox parameters to max out at 200px? Fyunck(click) (talk) 21:28, 19 April 2025 (UTC)[reply]
Ok, but the infobox discussion is one section above. Here I want to discuss the upright=1.35 size limit on non-infobox images. Should it be changed, and if so why? —David Eppstein (talk) 21:58, 19 April 2025 (UTC)[reply]
The default infobox image size is 250. Has been for a decade. It's right there in the policy. Magnolia677 (talk) 22:05, 19 April 2025 (UTC)[reply]
No. If you put no size parameter at all it has been 200 or 220 for years. Now it has suddenly ballooned to gargantuan size 250. And wiki has plenty of places that says the ideal size is 200px. Fyunck(click) (talk) 23:57, 19 April 2025 (UTC)[reply]
@Magnolia677: Which policy? Certainly not Wikipedia:Image use policy, because until a few hours ago, the figure 250 wasn't even mentioned. --Redrose64 🌹 (talk) 23:32, 19 April 2025 (UTC)[reply]
My error, it's 220 per Module:InfoboxImage. (the module uses 250 a lot). Magnolia677 (talk) 14:10, 20 April 2025 (UTC)[reply]
The 300px / |upright=1.35 thing was added in this edit, way back in July 2011 by Northamerica1000 (talk · contribs), but seems to have been undiscussed at the time. So: do we pick one of these to retain (which one), and adjust the other to suit? Using the text as it stood on 18 April 2025, this would give either of these:
  1. Therefore, it should be no wider than upright=1.2 (equivalent to 300px at the default preference selection of "250px"). ... Stand-alone lead images (not in an infobox) should also be no wider than upright=1.2.
  2. Therefore, it should be no wider than upright=1.35 (equivalent to 340px at the default preference selection of "250px"). ... Stand-alone lead images (not in an infobox) should also be no wider than upright=1.35.
My feeling is that retaining |upright=1.35 makes lead section images too much in-your-face, and that therefore we should go for my option 1. --Redrose64 🌹 (talk) 23:32, 19 April 2025 (UTC)[reply]
Perhaps because I run with default image size 300px and am used to it, I think retaining upright=1.35x to maintain proportionality with existing images in existing articles is a better choice. If an individual image is too in-your-face we can handle that on a case by case basis rather than insisting that all lead images be small even in cases where that might not be the right choice. Basically, I prefer more flexibility rather than more rigidity. Let's make a thought experiment: suppose the default image size were increased to 360px. Should we then demand that all lead images be reduced to upright=0.8? —David Eppstein (talk) 23:52, 19 April 2025 (UTC)[reply]
Noting that David Eppstein has already changed the page text to read "337px", which is neither of the options above. Re the thought experiment, we would not reduce the guidance number, because the change to larger thumbnails was caused by the increase in screen resolutions of almost all screens that are used to view Wikipedia. In a world where we would change the default to 360px, the number of pixels on a screen would have increased again. Our current situation, and the thought experiment, show why chasing or even mentioning px sizes is a fool's errand. All images displayed by templates (I'm looking at you, Module:Location map and Template:Infobox mapframe) should default to a fraction of the thumbnail preference. – Jonesey95 (talk) 20:55, 20 April 2025 (UTC)[reply]
If 337px is inaccurate, I would be happy for it to be changed. It is merely what you get when you multiply 1.35 (the status quo ante multiplier) with 250 (the new default pixel size). The important relevant part of this policy is the guidance on which multiplier to use for lead images; the corresponding pixel size is merely advisory. —David Eppstein (talk) 23:51, 20 April 2025 (UTC)[reply]
As I commented above, average screen size has increased substantially more over time than default thumbnail size. We should keep the previous values of upright - effectively increasing the pixel size by ~14% - to keep lede images being a useful size as screens continue to grow. Too-narrow images are particularly a problem in infoboxes - a very narrow infobox will end up unnecessarily long due to text wrapping. Pi.1415926535 (talk) 05:25, 21 April 2025 (UTC)[reply]
I don't know how that's possible. Sure, desktop screens are much bigger... I use one because of all the photo editing. But most people I know don't have desktops. They have tablets and 15-17 inch laptops. Some have nothing except their cellphones. Average screen size for most readers has gotten much smaller imho. Fyunck(click) (talk) 07:00, 21 April 2025 (UTC)[reply]
On my Android cellphone, at least, the Wikipedia app completely ignores the specified image size and shows the lead image full-width above all text of the article. So I think cellphone usage may be largely irrelevant for what we specify here. —David Eppstein (talk) 07:36, 21 April 2025 (UTC)[reply]
On an iphone it's not full width... it's massively zoomed in (often cutting off part of a persons scalp). But my point is, while screens started to get bigger and bigger, they have now gotten smaller and smaller for most folks. Fyunck(click) (talk) 08:17, 21 April 2025 (UTC)[reply]
The MediaWiki software rounds odd sizes up to the next multiple of 4; |upright=1.35 actually emits 340px for those with thumbnail set to 250px and all logged-out users. --Redrose64 🌹 (talk) 16:14, 21 April 2025 (UTC)[reply]
One thing to consider is that in the tech news that I linked in the previous section, it says File thumbnails are now stored in discrete sizes. If a page specifies a thumbnail size that's not among the standard sizes (20, 40, 60, 120, 250, 330, 500, 960), then MediaWiki will pick the closest larger thumbnail size but will tell the browser to downscale it to the requested size.. This means that if |upright=1.35 is specified, users will be served an image that is actually 500px wide, for downscaling to 340px client-side. But if |upright=1.2 is specified, users will be served an image that is actually 330px wide, for downscaling to 300px client-side. The scale factor here between 500 and 330 is 1.515, but the file size is more likely to be proportional to the number of pixels than to one dimension. Since the area factor is the square of the scale factor, this works out at 2.295. In other words, to retrieve an image for display at 340px the client is downloading about 230% more data than if it retrieves an image for display at 300px. --Redrose64 🌹 (talk) 08:34, 20 April 2025 (UTC)[reply]
That seems like something that could be adjusted very easily on the software end - if 340px ends up as a common thumbnail size, the developers could make it a standard size instead of 330px. Pi.1415926535 (talk) 18:06, 21 April 2025 (UTC)[reply]

I finally realized that the prior wording of this part of policy applied the same 1.35 multiplier maximum to both infoboxes and standalone lead images. While we sort this out I have temporarily reworded the policy to separate out these two different contexts. I support 1.35 for standalone lead images but I think it is far too wide for infoboxes. Was the old 0.9 multiplier for infoboxes intended as a default or as a max? If only the default, we should also have a max. Both 1 and 1.2 seem like reasonable possibilities to me. We might also suggest that images needing extra wideness (like panoramas) be taken out of the infobox. —David Eppstein (talk) 18:32, 21 April 2025 (UTC)[reply]