Commons talk:AI-generated media

From Wikimedia Commons, the free media repository
Jump to navigation Jump to search

Structured data suggestion[edit]

The policy might also want to mention something about how to model the structured data for these images, for instance by using P31 – instance of with Q96407327 – synthetic media (suggested in the Wikidata Telegram group). Ciell (talk) 11:59, 7 January 2023 (UTC)Reply[reply]

  • Is there also a way to record the generation prompts, and the generation engine / knowledgebase, in the structured data? Andy Dingley (talk) 12:57, 7 January 2023 (UTC)Reply[reply]
    No. You will need to propose said property on WD Trade (talk) 22:41, 20 August 2023 (UTC)Reply[reply]

Court cases[edit]

Hi, Getty Images is suing the creators of AI art tool Stable Diffusion for scraping its content in UK. And also Kelly McKernan, Karla Ortiz, and Sarah Andersen, three artists from the USA. Yann (talk) 12:56, 23 January 2023 (UTC)Reply[reply]

Artists Protest After ArtStation Features AI-Generated Images Hmm. I wonder if Commons are going to face similar backlash for allowing AI art Trade (talk) 13:54, 23 January 2023 (UTC)Reply[reply]
There are issues with the images from the protest being uploaded here. Commons:Deletion requests/File:AI protest image 4.jpg was an image based on an image from that protest which is a step removed. Ricky81682 (talk) 23:44, 22 February 2023 (UTC)Reply[reply]

Edit warring over DR examples[edit]

@Yann and Trade: Please discuss here instead of edit warring. Brianjd (talk) 10:37, 25 January 2023 (UTC)Reply[reply]

These are typical examples of images with issues, and we don't have so many examples right now, so it is useful to list them. Yann (talk) 12:11, 25 January 2023 (UTC)Reply[reply]
The examples are completely generic and barely shows any discussion. nor insight on the topic Trade (talk) 12:56, 25 January 2023 (UTC)Reply[reply]
Well, there are still useful as examples of files we do not want. If and when we have examples with more discussion, we could replace them (however this my never happen). Yann (talk) 13:24, 25 January 2023 (UTC)Reply[reply]
The disputed examples are:
I agree with Trade; we don’t need generic ‘personal artwork’ nominations here. They are completely useless as examples for users who cannot see the actual files.
Commons:Deletion requests/File:Algorithmically-generated portrait art of a young woman with long purple hair.png is not good either. It just talks about ‘bad taste’, without explanation; there is nothing in the filename that suggests bad taste. In fact, the linked Commons:Deletion requests/Algorithmically generated AI artwork in specific styles by User:Benlisquare suggests it might be in scope for its detailed description. It also refers to an unidentified talk page discussion. That would be two comments buried in the middle of en:User talk:Benlisquare#December 2022:
File:Algorithmically-generated portrait art of a young woman with long purple hair.png
(responding to ping) This is an AI image you made and uploaded to Commons, in your own words, generated with txt2img using the following prompts: Prompt: attractive young girl, large breasts.... Goodbye. Levivich (talk) 21:47, 23 December 2022 (UTC)Reply[reply]
Yes, and I agree this image was created in bad taste. However, I have not plastered it all over any Wikipedia articles. As a sign of good faith, I would have nommed it for speedy deletion out of compliance with the community, but I won't be able to do so due to the indef. --benlisquareTCE 21:56, 24 December 2022 (UTC)Reply[reply]
These comments must be read in the context of enormous controversy, on both projects, regarding that user’s uploads and the way that user was treated in response.
There are two lessons here:
  1. Remember that, once a file is deleted, non-admins can no longer see the file or its description and may not be able to see its usage. Ensure that deletion discussions include the required context, especially if they are to be used as precedents.
  2. Be careful which deletion requests you list as examples.
Brianjd (talk) 13:44, 25 January 2023 (UTC)Reply[reply]
That leaves Commons:Deletion requests/File:Algorithmically-generated portrait art of a young woman with long blonde hair.png (where I left a long comment that was never addressed) and Commons:Deletion requests/Files uploaded by 冷床系 (with some very questionable non-policy-based comments that don’t address the uploader’s arguments). Brianjd (talk) 13:52, 25 January 2023 (UTC)Reply[reply]
In #Purpose, Yann said: AI images can be accepted, but seeing that Commons could be easily overflooded with out-of-scope images from AI, we should restrict these to cases when they are really useful. Please point to an example of an AI image that can be accepted.
Wait, I just realised that’s not the point of that section. It’s called Out of scope concerns with AI images (emphasis added). No wonder 100% of the deletion requests listed as examples were closed as ‘delete’.
Let’s rename it to Scope concerns and find a more balanced list of examples. Brianjd (talk) 13:57, 25 January 2023 (UTC)Reply[reply]
The copyright section is just as bad: every example was closed as ‘delete’. We need more balance there too. Brianjd (talk) 13:57, 25 January 2023 (UTC)Reply[reply]
What would be balance? I added Commons:Deletion requests/File:A curly-haired indian woman with iridescent violet eyes holds a hand to her face.png for scope because it was a keep discussion. The issue seemingly was we don't have "any other images of India women with curly hair in this style" which I find ridiculous but whatever, that was the consensus. What example would have you of an copyright issue where people kept it anyways? -- Ricky81682 (talk) 06:24, 11 February 2023 (UTC)Reply[reply]
  • I don't think either of those are good examples. First, in neither case, it was no clear it was AI-generated at all. I added Commons:Deletion requests/Files uploaded by IAskWhatIsTrue as an attribution example. The debate was because the uploader kept fighting about whether "they" created it or an AI created it. If they did, they needed to follow VRT. In either case, the failure to make it clear how the artwork was created was the issue in the debate. -- Ricky81682 (talk) 06:24, 11 February 2023 (UTC)Reply[reply]

Metadata & reproducibility[edit]

Best practice for an educationally useful AI-generated image would probably involve including information on how exactly the image was generated. Which algorithm was used, on which set of training data, which query, etc. Can/should we standardize that somehow? El Grafo (talk) 10:59, 6 February 2023 (UTC)Reply[reply]

Agreed. Basic information like prompt, etc., should be mandatory. Yann (talk) 12:33, 6 February 2023 (UTC)Reply[reply]
I'm afraid mandatory might be a bit too much to ask, especially for imported files ... El Grafo (talk) 08:11, 7 February 2023 (UTC)Reply[reply]
I agree that "mandatory" goes too far. Ziko van Dijk (talk) 20:16, 8 February 2023 (UTC)Reply[reply]
It should be mandatory at least for images that the uploader claims as their own work, e.g. in order to enable the community to better evaluate that claim. That would address the issue mentioned by El Grafo. (Apropos, Ziko, is there a particular reason why you consistently withheld the prompt for the images at Category:AI images created by Ziko van Dijk?)
Regards, HaeB (talk) 07:39, 9 February 2023 (UTC)Reply[reply]
Mandatory for own work might work. On the other hand, we don't require source code for things like Category:Gnuplot_diagrams either, although it would make a lot of sense to do so. El Grafo (talk) 08:59, 9 February 2023 (UTC)Reply[reply]

"Special case: Deepfakes"[edit]

A deep fake is literally just any AI generated image in which a person have been replaced with someone else's likeness. Does it really need it's own policy? Trade (talk) 19:18, 10 February 2023 (UTC)Reply[reply]

  • I think it does. They are almost always unacceptable. - Jmabel ! talk 23:05, 10 February 2023 (UTC)Reply[reply]
    There is nothing in Commons:Photographs of identifiable people to suggest that one of type AI generated picture technique is worse than others.
    The section either needs to be renamed or reworked entirely Trade (talk) 21:40, 11 February 2023 (UTC)Reply[reply]

Intentionality[edit]

I disagree with the statements about intentionality introduced in these edits by User:Gnom. Yes, historically courts have ruled that copyright infringement requires intent, but we have no idea what they will rule about apparently derivative works by non-human intelligences. I could readily imagine a legal argument that the failure of the creators of the AI to take precautions against it creating infringing works is the sort of negligence that will be viewed just as much of an infringement as if there were intent. We just don't know. I believe that in discussing the legal situation at this time we need to consider this a gray area. - Jmabel ! talk 16:51, 18 February 2023 (UTC)Reply[reply]

Erm, at least under European copyright law, copyright infringement does not require intent. In other words, it doesn't matter whether someone intentionally or unintentionally uploads a derivative work on Commons, we still have to take it down. Gnom (talk) 19:13, 18 February 2023 (UTC)Reply[reply]

"This AI-generated image likely violates the copyright in Breakfast at Tiffany's"[edit]

The idea of displaying a AI generated copyvio image to illustrate the concept of copyvio is a bit weird isn't it? Trade (talk) 23:41, 18 February 2023 (UTC)Reply[reply]

Yes, but so far the image has not been deleted, for reasons unbeknownst to me. Gnom (talk) 00:02, 19 February 2023 (UTC)Reply[reply]
It’s a picture of Audrey Hepburn dressed like Holly Golightly. That’s not copyrightable, so this is a terrible example. I’m removing it and the other image because they don’t show how they’re doing so. It’d be better to use a fake or PD source to demonstrate the concept of derivative works. Obviously. Dronebogus (talk) 07:14, 19 February 2023 (UTC)Reply[reply]

Problems with rewrite[edit]

@Gnom: The recent rewrite introduced some problems, IMO. Primarily, it makes some bold statements implying that most AI generated images are derivative works of their training images. This is an oversimplification of the situation. An AI-generated image can be extremely derivative of one or more training images or it can be relatively original. This depends on a huge range of variables, including what type of training is used, how large the training set is, whether an image prompt is used, whether an artist prompt is used, how much weight those respective prompts are assigned, how much "chaos" the model adds or is instructed to add, etc. I think the wording needs a lot more nuance and should basically say that "it depends" rather than suggesting that most AI-generated images are copyvios (aside from fan art).

Another problem with the rewrite is that it suggests that a person must give their consent in order to be depicted in an image without violating their privacy. In the United States at least, that's a pretty alien concept and I think we would have a hard time gaining consensus for it on Commons. Personally, I agree with the statement, but I think most people on Commons don't agree with it, and if we want this page to have any chance of being enacted as a guideline, its initial wording should be conservative.

Finally, the claim that "the creators of image-producing algorithms may hold (co-)authorship in the output of their software" is simply not true and is contradicted by pretty much every source on the subject. You'll have to present some strong source citations to convince me that it's correct. Nosferattus (talk) 08:26, 19 February 2023 (UTC)Reply[reply]

Although there are now some problems with the current version. It claims, "In the United States and most other jurisdictions, only works by human authors qualify for copyright protection." which is true, but not helpful. The point is that in some jurisdictions there is copyright applicable to computer-generated works. So we can't simply write this off as a blanket statement. There will almost certainly need to be country-specific guidance here, such as we have to do for FoP.
Also the claim, "In 2022, the US Copyright Office confirmed that this means that AI-created artworks that lack human authorship are ineligible for copyright." keeps circulating. But this doesn't much apply to us either. Even if AI-creation doesn't engender a new copyright, there may still be an existing copyright involved (the DW situation). Most of the sensible discussion here on Commons has been concerned about that: a trained GAN AI system (with a large training dataset) appears likely to be judged for this quite differently to an automated system producing algorithmic or random art.
See also the Facebook thread, where a WMF staffer confidently asserts "Images made by non-humans are public domain." Andy Dingley (talk) 11:13, 19 February 2023 (UTC)Reply[reply]
I don't see the quoted comment starting with "Images made by non-humans" in the Facebook thread. Maybe the comment is hidden to the public, edited or inaccurately quoted? / Okay, I see it in one of the collapsed comments, but the person who said it doesn't seem to be a WMF staffer but a student. (I prefer not to name the student.) whym (talk) 12:00, 19 February 2023 (UTC)Reply[reply]
@Nosferattus: Can we please continue iterating the page from where I last left off? I really put a lot of thought into this – and maybe you allow me to mention that I have extensive academic and professional knowledge in this specific field. Gnom (talk) 14:43, 19 February 2023 (UTC)Reply[reply]
OK, some specific points:
  • I don't like the phrase "a certain likelihood". It's correct, but it's too likely to be read incorrectly as "a certainty", which is not what is meant. We don't need this phrase and we can use better. Even "likely" is clearer.
  • "When using prompts intended to imitate the style of a particular artist, there is a high risk that the algorithm uses copyrighted works by that artist to create the desired output. " I don't much like the phrase "high risk". We don't care about that. Our concerns should begin at a lower level, that of a "credible risk".
  • "it is likely that the [such] output constitutes a copyright infringement." This is true, but it's also ignoring any aspects of fair use or parody. Given the immature state of the law around AI generation, we don't know how those will be interpreted. Under US law (and in other sections we're taking a deliberately US-only view) these could be reasonable justification for AI use of them for training, without it then being seen as a violation. As none of us can cite an authoritative view on this as yet, we should describe the fuller situation.
  • "making them "unintentionally derivative works"" What's an "unintentionally derivative work"? (as distinct from a simple derivative work?) This is quoted as if that's a specific and distinct term, but I can find no source, definition or different legal treatment for it.
  • "Uploading AI-generated art on Commons may violate the rights of some algorithm providers" What is the basis for this? Under what jurisdictions would such rights be enforceable? (I know of none). There is a difference between a computer system providing some output, which is then the copyright material of the system's owner / operator (UK law at least recognises this concept) and the provider of an AI algorithm or system then owning rights to the copyright of any material produced by it in the future. That's a step further and I know of no jurisdiction that would recognise such.
Andy Dingley (talk) 17:05, 19 February 2023 (UTC)Reply[reply]
Per Gnom's request, I restored the rewrite and instead edited it to address some of the concerns raised above. I still have strong concern about the privacy wording, but I'm not sure what it should say instead. I basically agree with Gnom's wording there, but I don't think it reflects consensus on Commons. Nosferattus (talk) 17:53, 19 February 2023 (UTC)Reply[reply]
Thank you! I will look into the various points and probably respond tomorrow. Gnom (talk) 18:04, 19 February 2023 (UTC)Reply[reply]
Regarding "Uploading AI-generated art on Commons may violate the rights of some algorithm providers", note that this unsubstantiated legal claim (which had beenadded together with the "However, most algorithm providers renounce all rights ..." statement that is quite clearly false, see below) has since already been changed back by Nosferattus to refer to terms of use as before the rewrite (current wording: Uploading AI-generated art on Commons may violate the terms of use of some AI providers). In any case, the following sentences clarify that the Commons community does not mere TOU violations (that are not also copyright violations) as grounds for deletion, per COM:NCR. Regards, HaeB (talk) 16:12, 20 February 2023 (UTC)Reply[reply]
I agree with many of Nosferattus' concerns. In particular, Gnom has made several bold claims about various frequencies that I find highly dubious (based on what I have read so far in various sources that I consider fairly reliable about this area of machine learning and/or copyright, happy to spend time to go into more detail if Gnom is willing to divulge his evidence first). Besides those about the probabilities of copyright violations in particular situations, this also includes Gnom's claim (since removed by someone else) that "most algorithm providers renounce all rights in the output through their terms of service." Based on a quick check, that's evidently not true for Midjourney and Craiyon. It also wasn't true for DALL-E at the time of this deletion discussion.
Regarding "I have extensive academic and professional knowledge in this specific field": I am willing to assume that, given his education and current profession, Gnom has good professional knowledge of at least German Urheberrecht black-letter law. But that doesn't mean that the Commons community is required to espouse his personal theories about a new and currently widely debated and contested legal topic, especially if he isn't even willing to share citations or other evidence (despite this being the norm in many academic or professional legal contexts). Also, while I don't doubt Gnom's longtime personal dedication to the Wikimedia movement's mission, I would also encourage some reflectiveness about potential conflicts of interest (or at least conflicts of perspective) in cases where one's job regularly involves taking the side of companies and individual professionals against alleged violators of their intellectual property.
Regards, HaeB (talk) 14:37, 20 February 2023 (UTC)Reply[reply]
It seems that the likelihood of an AI work being a derivative also depends highly on the AI model being used. According to this paper, DALL-E is poor at mimicking artist styles, while Stable Diffusion is very good at it. Basically, I think there are too many variables to draw any sweeping conclusions. In most cases, the derivative issue will need to be handled on a case-by-case basis (given the absence of any clear legal precedence, which may change after the Warhol Foundation v. Goldsmith decision comes out). Nosferattus (talk) 19:53, 20 February 2023 (UTC)Reply[reply]
A good point in general about differences between AI models; but even so note that artist styles are not copyright protected, see e.g [1] or [2]. This fact is well-known among many artists - but apparently not to those who are recently engaging in this kind of "AI stole my style" hue and cry. Regards, HaeB (talk) 00:05, 21 February 2023 (UTC)Reply[reply]
That's also a good point that might be worth mentioning in the derivatives section. Nosferattus (talk) 01:27, 21 February 2023 (UTC)Reply[reply]
The problem is that considering the amount of AI-generated media that is currently being uploaded on Commons, there is a significant risk of inadvertent (and near-impossible to detect) copyright and privacy violations being among them. And we need to clearly acknowledge this risk on this page. Once we have done that, we need to have a discussion about what to do with this risk. Ignore it? Or ban all AI-generated media altogether? Or some other solution? With the current hype, it's probably difficult to have a proper discussion about the latter for the time being, but at's least get the first part right. --Gnom (talk) 22:54, 22 February 2023 (UTC)Reply[reply]
Keep in mind that currently this page is not an official guideline or policy, so even the very basic guidance we have outlined so far is not official or actionable. If we hope to have the community enact this as a guideline, its initial version should be something that nearly everyone on Commons can agree with. In other words, its initial wording needs to be rather "soft" (except where there is clear consensus or legal precedent to lay down harder rules). It is extremely difficult to get the community to adopt a new set of official guidelines. Please don't underestimate this. If the wording is even slightly controversial, there is no chance it will achieve consensus. After the initial guidelines are adopted, there will be plenty of opportunity to discuss tightening them and responding to changes in the legal landscape. Nosferattus (talk) 01:08, 24 February 2023 (UTC)Reply[reply]
Last month, the U.S. copyright office again asserted that AI generated works have no copyright protection (as original works): https://www.copyright.gov/docs/zarya-of-the-dawn.pdf. This is the third time that the U.S. Copyright office has made such an assessment. Thus the theory that AI providers have a copyright stake in AI generated works is bogus. Even the UK, which provides a limited copyright protection for the prompter, provides no copyright ownership for the software developers or providers. I'm going to revert the changes that were made regarding this. Nosferattus (talk) 17:43, 18 March 2023 (UTC)Reply[reply]

Privacy of living people[edit]

Starting a new section to specifically discuss the "Privacy of living people" section. I don't think the current wording reflects the consensus on Commons, but I'm not sure what to change it to. Suggestions? Nosferattus (talk) 17:55, 19 February 2023 (UTC)Reply[reply]

At the very least, the claim that "there is a certain likelihood that any AI-generated image with a human face in it violates the privacy of a living person" should be backed up by evidence or removed (also given that other hand-wavy probability claims added by the same user turn out to be highly dubious upon closer inspection, see above). Regards, HaeB (talk) 15:16, 20 February 2023 (UTC)Reply[reply]
I am concerned about scenarios like these:
Gnom (talk) 22:19, 22 February 2023 (UTC)Reply[reply]
Just to remind everyone: these might be issues in some countries, but basically are not issues in the U.S. (where our servers are hosted) unless the image is somehow disparaging or is used in a way that implies an endorsement by that person. - Jmabel ! talk 22:39, 22 February 2023 (UTC)Reply[reply]
@Gnom: The section that discusses privacy issues is under the question "Can AI-generated media be lawfully hosted on Commons?". Can you elaborate on what laws you are concerned with? While I share your general concern for the privacy of living people, we need to focus on what the actual legal issues are (and how Commons typically deals with those legal issues). If your concerns are not legal in nature, they should be moved to a different section. Nosferattus (talk) 00:49, 24 February 2023 (UTC)Reply[reply]
On a legal level, this is of course a violation of the GDPR insofar as EU residents are being depicted. Under U.S. law, I am also quite confident that would be illegal to host such images, but others are probably in a better position to assess this. On a non-legal level, I can only say that I would go berserk if an AI were to spit out my face under certain conditions. Gnom (talk) 08:12, 24 February 2023 (UTC)Reply[reply]
@Gnom AI-generated art may show faces of living people who were not asked for their consent to appear in the output because photos or drawings depicting them were used to train the software. Accordingly, there is a certain likelihood that any AI-generated image with a human face in it violates the privacy of a living person. Under the Precautionary Principle, these respective files may be nominated for deletion.
No. There are a significant number of countries where consent is not required, and Commons generally does not impose its own consent requirements unless the photo seems to be undignified.
Ironically, the page includes an AI-generated image with a human face in it (Mona Lisa by Midjourney AI.jpg). Should it be deleted under the precautionary principle? Of course not, but there’s nothing on the page to explain this.
Where we can re-use existing laws, guidelines and policies, we should do so. This page should cover only AI-specific issues, and link to other pages for the other issues. Brianjd (talk) 15:02, 25 February 2023 (UTC)Reply[reply]
In the US (where Commons is hosted), Australia (where I live), and some other countries, it is generally legal to take and publish images of people in public without their consent. The EU and the US have very different attitudes to privacy, and people living in one of these jurisdictions must be very careful to avoid false assumptions about the other. Brianjd (talk) 15:05, 25 February 2023 (UTC)Reply[reply]
On another matter, why focus on faces? Images of people can violate privacy without depicting faces. Brianjd (talk) 15:06, 25 February 2023 (UTC)Reply[reply]
Hi, I agree that the Mona Lisa is a bad example (and I should note that was not the one to put it there). I would prefer the examples I inserted above instead. The left image, assuming(!) it were to actually show the face of a real underage girl, would give her parents a right to sue the Wikimedia Foundation for violating their daughter's privacy, even under U.S. law.
Also, you are of course correct that not only the depiction of a person's face can constitute a violation of their privacy, but that is where I see the main issue. Gnom (talk) 20:17, 4 March 2023 (UTC)Reply[reply]
There is a problem if an image shows an actual person by the actual original image. But... when the AI creates a new image based on a dozen or thousands of original images, and someone says: "That new image looks like person X!", is that a privacy issue? How decides that the image looks like person X? How to determine that; "looking alike" can be the subjective judgement of an individual. I for example believe that Tim Curry looks like Edmond Privat. Ziko van Dijk (talk) 18:59, 5 March 2023 (UTC)Reply[reply]
I would think this is an area where we would not want to host things that are in the gray zones. - Jmabel ! talk 20:49, 5 March 2023 (UTC)Reply[reply]
@Ziko The cases that I am thinking about are those where there is simply no question whether or not the AI-generated image depicts a real person. Gnom (talk) 08:08, 6 March 2023 (UTC)Reply[reply]
@Gnom, can you give an example of what you mean? In which context would there be no question? Ziko van Dijk (talk) 08:13, 6 March 2023 (UTC)Reply[reply]
If the similarity between the AI-generated human face and their real likeness is so close that if you asked 100 people, 99 would say that the two images (the AI-generated image and a photo of the actual human face) depict the same person. That's the scenario I would like to discuss. Gnom (talk) 15:04, 6 March 2023 (UTC)Reply[reply]
You vastly overestimate how skillet your average person are at telling the difference Trade (talk) 21:49, 18 March 2023 (UTC)Reply[reply]

@Gnom: While I agree with you in principle, I think you have a misunderstanding of U.S. law. In the U.S. there are very very few protections for privacy and strong protections for expression. For example, in the U.S. I could go to a public playground, take photos of a 5-year-old child, ask no one's permission or consent, and then post those photos on the internet or in an art show or whatever. Not only that, but I could even take photos of someone through the window of their home, and in many states it would be perfectly legal. As long as there is no nudity involved (which triggers the federal video voyeurism law) or commercial use involved (which triggers rights of publicity), you can violate people's privacy all day long in the U.S. If you don't believe me, read [3]. This is what happens when you enshrine free speech in your constitution, but not privacy. This lack of privacy protection is also what specifically caused Roe v. Wade to recently be overturned in the U.S. Supreme Court. The Court ruled that because there is no actual "right to privacy" in the U.S., Roe v. Wade was invalid and thus abortion can be prohibited by the law. Of course, I don't believe that Commons policy on privacy should be based on whatever is allowed by U.S. law, but we can't pretend that such a policy is based on the law when it isn't. Our policy should be based on basic ethics and respect for human dignity. But we have to convince the Commons community to agree to such principles in this case. It can't just be taken for granted. Nosferattus (talk) 19:38, 18 March 2023 (UTC)Reply[reply]

Since there doesn't seem to be consensus for the contents of this section, I've removed it. I would ideally like to replace it with some sort of guidance we can agree on. Does anyone have suggestions? Nosferattus (talk) 18:46, 19 August 2023 (UTC)Reply[reply]

AI-improved/restored photographs[edit]

Some software can automatically:

  • "Improve" a picture of en:Lake Titicaca using generative AI to transform green areas into realistic trees, transform dots into en:Florida ducks, etc.
  • "Restore" an old black&white picture of en:Léon Blum and colorize it, split hair into individual strands of hair, transform a photo grain fleck into a realistic en:wart somewhere on the skin, etc.

The resulting pictures are works of art/hallucination more than encyclopedic representations of reality. Do we already have a policy page for these? If not, should we draft a policy here?

By the way, I recently raised a concern about the rising trend of camera apps to perform such "improvements" by default without the user even knowing, it might be too early to worry but here it is for possible reference: Commons:Village_pump#Mainstream_cameras_generating_non-existent_artifacts_(AI_generative_upscaling?),_what_can_we_do_about_it?

Thanks! Syced (talk) 06:36, 13 March 2023 (UTC)Reply[reply]

"Artificially upscaling or enlarging using any tool, including AI-based or deep learning services" get a mention at COM:OVERWRITE, although only to say that such versions shouldn't overwrite the original. I have sadly seen a few cases of this recently, including bad MyHeritage output that just fakes extra detail onto the eyes, nose and mouth and makes sometimes absurd guesses about facial hair.
Commons:Deletion requests/Files uploaded by Madelgarius is an open discussion about whether to delete a hundred such upscaled portrait photos from Commons. Belbury (talk) 13:26, 13 March 2023 (UTC)Reply[reply]
@Syced: Other than the guidance already provided at Commons:AI-generated media#Categorization, I think someone should create a warning template for such cases that can be applied on the File pages. Nosferattus (talk) 18:37, 18 March 2023 (UTC)Reply[reply]

Uploads by PixelPenguin87[edit]

PixelPenguin87 (talk · contribs)

Might wanna contact the legal team on what to do with these type of AI generated images. This sort of photorealistic AI generated content could potentially be a violation of Commons:CHILDPROTECT Trade (talk) 21:46, 18 March 2023 (UTC)Reply[reply]

@Brianjd, Ricky81682, and Nosferattus: --Trade (talk) 21:53, 18 March 2023 (UTC)Reply[reply]
@King of Hearts: --Trade (talk) 03:28, 19 March 2023 (UTC)Reply[reply]
I fully agree. A WMF Legal clarification would be useful. Wutsje 23:06, 18 March 2023 (UTC)Reply[reply]
See also U93rFh2T (talk · contribs), VibrantExplorer372 (talk · contribs) and BlueMoon2023 (talk · contribs). Wutsje 03:08, 19 March 2023 (UTC)Reply[reply]
Have you tried to seek WMF out for a legal clarification? Trade (talk) 15:49, 21 March 2023 (UTC)Reply[reply]
I don't think we need legal advice to decide just to delete this sort of thing. - Jmabel ! talk 16:35, 21 March 2023 (UTC)Reply[reply]
I didn't wanted to risk going against community consensus. Trade (talk) 20:16, 31 March 2023 (UTC)Reply[reply]

Hosting of AI-generated media should follow a whitelist model[edit]

that is, uploads are allowed only with prior approval by the community. in some way similar to com:bots, which must be approved before deployment.

the reason is, ai is not only capable of generating images but realistic audio and video. this page and the discussions so far have only considered images. only 1 section above mentions deepfakes, which are one kind of ai videos. ai can do videos much better than deepfakes. ai can also do audio: https://ai.googleblog.com/2022/10/audiolm-language-modeling-approach-to.html https://google-research.github.io/seanet/audiolm/examples/ .

i believe it's not beneficial or ethical to host ai-generated media alongside manual creative works. even if it's allowed, ai-generated media should be clearly tagged as such, and put into a separate category. for example, ai-generated media about cities -- be it images or audiovisual files -- should have its own "category:AI-generated media about cities" under Category:Cities. ai photos should not be mixed with human-made photos. ai audio should not be mixed into Category:Audio files of cities. ai videos should not be mixed with other human-made videos in Category:Videos by city.

we should use a whitelist model so as to deter any unwanted ai-generated media before the situation becomes unmanageable. RZuo (talk) 17:00, 21 March 2023 (UTC)Reply[reply]

Hello @RZuo, I agree that AI content should be made identifiable as such, in a suitable way. However, I disagree that all uploads need a prior permission by the community. This would also be difficult, because, how should community members know the image before it is uploaded? "The community" might be only a handfull of people who are really so committed to the cause that they would invest hours in related discussions. / I am not quite sure what do you mean by a whitelist in this context, but maybe I did not understand correctly. Kind regards, Ziko van Dijk (talk) 16:15, 22 March 2023 (UTC)Reply[reply]
whitelist.
better deny any upload instead of sieving uploaded files for ai content. it's impossible to do the latter, especially with audio. photos are also super realistic now.
anyone uploading ai content without approval should be indef blocked on sight and their uploads deleted. RZuo (talk) 16:24, 22 March 2023 (UTC)Reply[reply]

m:Wikilegal/Copyright Analysis of ChatGPT has been published a couple of days ago. It is primarily about text, but it also briefly mentions AI-generated images. whym (talk) 13:54, 24 March 2023 (UTC)Reply[reply]

United Kingdom[edit]

{{PD-algorithm}} contains a note "The United Kingdom provides a limited term of copyright protection for computer-generated works of 50 years from creation", citing this UK Intellectual Property Office page. From this page:

Unlike most other countries, the UK protects computer-generated works which do not have a human creator (s178 CDPA). The law designates the author of such a work as “the person by whom the arrangements necessary for the creation of the work are undertaken” (s9(3) CDPA). Protection lasts for 50 years from the date the work is made (s12(7) CDPA).

I think it might be tricky to ascertain who "the person by whom the arrangements necessary for the creation of the work are undertaken" might be in modern AI generated media (the author of the prompt? The programmer(s) of the software? Both? - because the software itself is certainly "necessary for the creation of the work", and the prompt as well), but it seems that in the UK, AI generated media is protected anyway, even if we might be unsure who's the owner of the rights. The Office also states there:

The UK remains one of only a handful of countries worldwide that provides this protection. Investment in AI has taken place in other countries, such as the United States, which do not provide this type of protection. Some people argue that this protection is not needed, and others that it should be provided differently.

So, I think that this page should be amended, too, in some way, to contain this information that in the UK and in a "handful of countries" (which countries?) there is protection for AI art, and in deletion discussions, I assume that we have to check where the image was generated - if generated in the US or most other countries, it's {{PD-algorithm}}, but if generated in the UK or the "handful of countries" with protection, we must delete it. In the case of Commons:Deletion requests/File:Alice and Sparkle cover.jpg which I decided to keep, I think it's fine, as the author of the prompt per this article is currently living in California, and was using Midjourney, which is based in San Francisco. Gestumblindi (talk) 20:58, 9 June 2023 (UTC)Reply[reply]

It's possible the author didn't knew what other countries he was referring too and simply made a reasonable assumption Trade (talk) 21:05, 9 June 2023 (UTC)Reply[reply]
Well, to remedy the lack of UK information, I made an addition based on this UK Intellectual Property Office information. Gestumblindi (talk) 18:24, 10 June 2023 (UTC)Reply[reply]

A large-scale AI-generated images upload[edit]

Just for the records, in Commons:Deletion requests/Files in Category:AI images created by David S. Soriano I would have decided exactly like Minorax. Per COM:INUSE, it would be hard to argue for deleting files that are actually already in use - and they were added to articles and user pages by various users, not by the uploader -, and this shows that, apparently, these images can be considered useful, but on the other hand, I think there is really a "slippery slope": If we just blanket kept the images, the door would be open for flooding Commons with thousands, tens of thousands - well, why not millions? - of such images, as it's very easy to mass generate this kind of content. Gestumblindi (talk) 09:11, 18 June 2023 (UTC)Reply[reply]

Can we know how stable those usages are, beyond the binary in-use vs not in-use? If it has to be checked manually, I hope there is a better tooling to track usage over time. It looks like File:UFO From Distant Past.png was kept for being in use, but it's not in use at the moment. whym (talk) 03:15, 19 June 2023 (UTC)Reply[reply]
Given that the remaining files appear mostly on user pages, I'd expect that any mainspace project usage was being washed out days or weeks later. File:UFO From Distant Past.png is the seventh result in a Commons search for "ufo" and certainly the most striking and exciting depiction, but if anyone added it to a project article about UFOs it would (as an illustration of no particular UFO case, drawn in part or in full by a hallucinating AI) be reverted or replaced with something more useful. The same goes for Soriano's cubism uploads: a superficially attentive editor might search Commons for "cubism" and add one to an article as an example of that style, but an editor paying more attention would replace it with a free image by a notable and human artist. Belbury (talk) 10:59, 19 June 2023 (UTC)Reply[reply]
This would be far easier to solve if anyone could come in contact with David Trade (talk) 01:38, 27 June 2023 (UTC)Reply[reply]

Can the prompts themself be copyrighted?[edit]

I am seeing a trend where users have two separate copyright tags. One for the AI prompt they used and one for the AI generated image that came from the output. Should we allow this> Trade (talk) 21:05, 28 June 2023 (UTC)Reply[reply]

And plenty of others i cant currently remember. Trade (talk) 01:21, 29 June 2023 (UTC)Reply[reply]

Those are probably on the borderline of copyrightability (the first more likely than the second, in my view). Given that in both cases they are specifically disavowing copyright on the prompt and placing it in the public domain, it's hard to see anything objectionable. I think I'd object to someone putting something here, claiming copyright on their prompt, and insisting that using even the prompt requires conforming to a license. - Jmabel ! talk 02:00, 29 June 2023 (UTC)Reply[reply]
  • Are there any of these where the prompt is being released under a licence? The examples here are really the opposite of this: the uploader has put the prompt deliberately into [sic] the public domain, to clearly disavow any licensing. Andy Dingley (talk) 02:05, 29 June 2023 (UTC)Reply[reply]
I think it's fine if the uploader is explicitly declaring the prompt public domain (or CC0). Other cases would need a closer look. Nosferattus (talk) 01:50, 16 July 2023 (UTC)Reply[reply]
If the author have the right to license a prompt as PD then they have the right to license the prompt under as CC-BY-SA. Allowing the former but not the latter is essentially saying that they only own the copyright to the prompt as long as it is convenient for Commons Trade (talk) 02:12, 16 July 2023 (UTC)Reply[reply]

Prompts need to be disclosed[edit]

It seems to me that the main way AI-generated images are of use is precisely as examples of AI-generated images. To that end, I propose that we should (going forward) require that the file page text for all AI-generated images include, at a minimum:

  1. what AI software was used
  2. what prompt was used

I have no problem with some images being grandfathered in, and of course exceptions need to be made for images that are, for one or another reason, notable in their own right. There might even be a reason for some other exception I'm not thinking of, but I've been seeing floods of low-value AI-generated images, often not marked as such, often involving sexualized images of young women or girls. If those last are coming from prompts that do not ask for anything of the sort, then that itself would be important to document. If, as I suspect, they come from prompts that ask for exactly that, then they really don't belong here, any more than someone's personal photos of themselves and their friends.

And, yes, I am discussing two separate (but related) issues here, disclosure and scope. - Jmabel ! talk 16:53, 28 July 2023 (UTC)Reply[reply]

For a starter, we need a category for AI images without prompts so we can keep track of the issue Trade (talk) 18:56, 19 August 2023 (UTC)Reply[reply]
(which Trade introduced with Category:AI images generated using unidentified prompts). - Jmabel ! talk 23:55, 19 August 2023 (UTC)Reply[reply]

Question about files that are most likely AI-generated, but without definitive proof[edit]

Hello, I was wondering what best practice would be in cases where a file is found online that looks very clearly like an AI-generated image, but the creator either did not specify that they were created with AI or they claimed it to be their own work. Examples include this film poster (notice the pupils and hair), as well as this poster (notice the inconsistent teeth, the bizarre foot, the inconsistent windows on buildings, and the weird birds in the sky). Should we allow them to be uploaded under the presumption that they are AI, or should we assume that they're copyrighted unless they're explicitly stated to be AI? Di (they-them) (talk) 00:16, 20 August 2023 (UTC)Reply[reply]

Even if an image contains a lot of obvious AI artefacts, you'd also have to assume that there was no subsequent creative human editing of that image, which seems unknowable. Belbury (talk) 08:02, 20 August 2023 (UTC)Reply[reply]
There is exactly zero benefit from trying to do original research on whether or not a image was AI generated without the author admitting so. You are just asking to create a huge copyright mess Trade (talk) 22:38, 20 August 2023 (UTC)Reply[reply]

Is the note unnecessary?[edit]

The note "Although most current media-generating programs qualify as machine learning and not true artificial intelligence, the term 'artificial intelligence' is commonly used colloquially to describe them, and as such is the term used on this page" seems unnecessary to me. Media-generating programs are indeed artificial intelligence as well as machine learning. Machine learning is considered a subset of artificial intelligence. Artificial intelligence doesn't only refer to LLMs, and includes DALLE-2, Midjourney, etc. Machine learning might be more precise, but AI isn't incorrect. I would like to hear what other people think about this. Chamaemelum (talk) 18:32, 20 August 2023 (UTC)Reply[reply]

I agree with your assessment. Machine learning is a subset of AI. Nosferattus (talk) 19:49, 20 August 2023 (UTC)Reply[reply]

Real-life[edit]

@Trade: you added the parenthetical phrase in "AI fan art of (real-life) fictional characters", which seems oxymoronic to me. How can something be both "real-life" and "fictional"? - Jmabel ! talk 01:52, 21 August 2023 (UTC)Reply[reply]

Fictional characters that exist outside of the AI generated art in question Trade (talk) 01:54, 21 August 2023 (UTC)Reply[reply]
I hoped the name of the category was enough but unfortunately people keep filling it with images that had nothing to do with fan art. Trade (talk) 01:56, 21 August 2023 (UTC)Reply[reply]
To be fair, the description at the top of Category:AI-generated fictional characters doesn't suggest not to. And a lot of categorisation happens based on the category name alone.
Would Category:AI-generated fan art be a useful subcategory to create? Belbury (talk) 09:13, 21 August 2023 (UTC)Reply[reply]

AI images of real subjects (aka. "deepfakes")[edit]

One subject which this draft doesn't seem to address clearly is the topic of AI images which appear to represent real subjects - e.g. real people, real places, real historical events, etc. These images have the potential to be misleading to viewers, and can cause harm to the project by discouraging the contribution of real images, or by being used as substitutes for real images which are already available.

I'd like to address this as follows. This is intentionally strongly worded, but I feel that it's warranted given the potential for deception:

AI-generated images which contain photorealistic depictions of notable people, places, or historical events have the potential to deceive viewers, and must not be uploaded.

If AI-generated images containing these subjects are used as illustrations, effort should be made to use images which cannot be mistaken for photographs, e.g. by prompting the image generation model to use a cartoon art style.

In a limited number of cases, realistic images containing these subjects may be used as demonstrations of AI image generation or "deepfakes". These images should be watermarked to make it clear to viewers and downstream users of these images that they were machine-generated.

Thoughts? Omphalographer (talk) 23:23, 11 September 2023 (UTC)Reply[reply]

A COM:WATERMARK on a demonstration image significantly reduces any constructive reuse of it. Anybody wanting to reuse a notable fake like File:Pope Francis in puffy winter jacket.jpg in their classroom or book should be able to get that direct from Commons.
These images would benefit from prominent warning templates, though, and perhaps an explicit "Fake image of..." in the filenames. Belbury (talk) 08:25, 12 September 2023 (UTC)Reply[reply]
Why not just add a parameter to the AI template that can be used to indicate whether or not the image depicts a living person? Trade (talk) 10:51, 12 September 2023 (UTC)Reply[reply]
The problem I'm concerned with is reuse of these images outside Wikimedia projects, where the image description certainly won't be available and the filename will likely be lost as well. Photorealistic AI-generated images of recognizable subjects should be fairly rare on Wikimedia projects, and I'm confident that editors can come up with some way of marking them which makes their nature clear without being overly intrusive.
How about the rest? Are we on board with the overall principle? Omphalographer (talk) 19:41, 12 September 2023 (UTC)Reply[reply]
Seems reasonable to me. And an alternative to a watermark in the narrow sense would be a mandatory notice in a border under the photo. - Jmabel ! talk 20:10, 12 September 2023 (UTC)Reply[reply]
No matter the amount of whistles, alarms and whatnot you put up there will always be someone who cant be bothered to read it before posting the image somewhere else Trade (talk) 15:41, 14 September 2023 (UTC)Reply[reply]
Certainly. But if the image itself can tell viewers "hey, I'm not real", then at least it has less potential to mislead. Omphalographer (talk) 17:01, 14 September 2023 (UTC)Reply[reply]
In what manner is that not covered by the current AI-image license template + related categories and description? Trade (talk) 22:41, 23 September 2023 (UTC)Reply[reply]
Because those do not tend to travel with the image itself when it is reproduced. Indeed, if it the image is used incorrectly within a Wikipedia there would be no indication of that unless someone clicks through. - Jmabel ! talk 22:51, 23 September 2023 (UTC)Reply[reply]
Even if you click to expand an embedded image in an article, the full license and disclaimer templates are only visible if you then click through to the full image page. To an unsophisticated user, they might as well not exist. Omphalographer (talk) 00:19, 24 September 2023 (UTC)Reply[reply]
So in short the only realistic solution would be a warning template that appears when someone from a Wiki project click to expand the image Trade (talk) 22:04, 13 October 2023 (UTC)Reply[reply]
@Omphalographer: I think we should change "notable people" to "actual people" and remove "places" (as that seems overly broad and unnecessary to me). Nosferattus (talk) 23:14, 20 September 2023 (UTC)Reply[reply]
We may also want to clarify that this doesn't apply to AI-enhanced photographs. Nosferattus (talk) 23:16, 20 September 2023 (UTC)Reply[reply]
Excellent point on "notable people"; I agree that this policy should extend to any actual person, not just ones who cross some threshold of notability.
The inclusion of "places" was intentional. A synthetic photo of a specific place can be just as misleading as one of a person or event; consider a synthesized photo of a culturally significant location like the Notre-Dame de Paris or Mecca, for instance.
AI-enhanced photographs are... complicated. There's no obvious line dividing photos which are merely "AI-enhanced" and ones which begin to incorporate content which wasn't present in the source photo. For instance, the "Space Zoom" feature of some Samsung phones replaced photos of the moon with reference photos of the moon - this level of processing would probably be inappropriate for Commons photos. Omphalographer (talk) 00:11, 21 September 2023 (UTC)Reply[reply]
@Omphalographer I think there are some legitimate reasons for creating and uploading photorealistic AI images of places, and less danger that they cause harm. For example, an AI generated image of an ice-free Greenland might be useful for a Wikibook discussing climate change. Sure, it could be misleading if used in the wrong context, but it doesn't worry me as much as AI images of people.
So are you suggesting that all AI-enhanced photographs should also be banned? This will probably be the majority of all photographs in the near future, so I wouldn't support that proposal. Nosferattus (talk) 00:32, 20 October 2023 (UTC)Reply[reply]
I'm not suggesting that all AI-enhanced photos should be banned, but that the limits of what's considered acceptable "enhancement" need to be examined. Filters which make mild changes like synthetically blurring the background behind a person's face or adjusting contrast are almost certainly fine; ones which add/remove substantial elements to an image or otherwise dramatically modify the nature of the photo (like style-transferring a photograph into a painting or vice versa) are probably not.
With regard to "places", would you be happier if that were worded as "landmarks"? What I had in mind was synthetic photos of notable buildings, monuments, or similarly specific places - not just any location. Omphalographer (talk) 05:57, 20 October 2023 (UTC)Reply[reply]
 Oppose - Very strongly against this proposal, which would be highly problematic for many reasons and unwarranted censorship.
Agree with Belbury on prominent warning templates, though, and perhaps an explicit "Fake image of..." in the filenames - we should have prominent templates for AI images in general and prominent warning templates for deepfake ones...a policy on file title requirements is something to consider. Prototyperspective (talk) 22:11, 13 October 2023 (UTC)Reply[reply]
 Oppose per Prototyperspective. I fail to see why this issue is fundamentally different from other kinds of images that could be convincingly misrepresented as actual, unaltered photographs depicting real people, real places, real historical events, an issue that is at least a century old (see e.g. w:Censorship of images in the Soviet Union, Category:Manipulated photographs etc).
That said, I would support strengthening existing policies against image descriptions (and files names) that misrepresent such images as as actual photos, whether they are AI-generated, photoshopped (in the sense of edits that go beyond mere aesthetics and change what a general viewer may infer from the image about the depicted person, place etc.) or otherwise altered. That's assuming that we have such policies already - do we? (not seeing anything at Template:Commons policies and guidelines)
PS regarding landmarks: I seem to recall that authenticity issues have been repeatedly debated, years ago already, in context of Wiki Loves Monuments and related contests, with some contributors arguing that alterations like removing a powerline or such that "ruins" a beautiful shot of a monument should not affect eligibility. I do find that problematic too and would support at least a requirement to clearly document such alterations in the file description.
Regards, HaeB (talk) 01:35, 25 October 2023 (UTC)Reply[reply]
 Weak oppose per HaeB. Although I'm sympathetic to the idea of banning deepfake images, I think the proposed wording is too broad in one sense (subjects included) and too narrow in another sense (only addressing AI images). I would be open to a proposal focusing on photo-realistic images of people or events that seem intended to deceive or mislead (regardless of whether they are AI generated or not). Nosferattus (talk) 04:47, 25 October 2023 (UTC)Reply[reply]

Custom template for upscaled images[edit]

This page currently advises adding {{Retouched picture}} to upscaled images, which if used without inserting specific text gives a neutral message of This is a retouched picture, which means that it has been digitally altered from its original version. with no mention of the AI nature of the manipulation.

Would it be useful to have a custom AI-upscale template that puts the image into a relevant category and also spells out some of the issues with AI upscaling (potentially introducing details which may not be present at all in the original, copyrighted elements, etc), the way that {{Colorized}} specifically warns the user that the coloring is speculative and may differ significantly from the real colors? Belbury (talk) 08:19, 4 October 2023 (UTC)Reply[reply]

Prototyperspective (talk) 09:44, 5 October 2023 (UTC)Reply[reply]

I've made a rough first draft of such a template at {{AI upscaled}}, which currently looks like this:

This image has been digitally upscaled using AI software.

This process may have introduced inaccurate, speculative details not present in the original picture. The image may also contain copyrightable elements of training data.

When the template is included on a file page it adds that file to Category:Photos modified by AI per the recommendation at Commons:AI-generated_media#Categorization_and_templates.

Feedback appreciated on what the message should say, and what options the template should take. It should probably always include a thumbnail link to the original image (or an alert that the original is freely licenced but hasn't been uploaded to Commons), and an option to say what software was used, if known, so that the file can be subcategorised appropriately.

It may well be worth expanding this to a generic AI template that also covers restoration and generation, but I'll put this forward for now. --Belbury (talk) 08:21, 12 October 2023 (UTC)Reply[reply]

Could you make a template for AI misgeneration? Trade (talk) 23:41, 18 November 2023 (UTC)Reply[reply]
Would that be meaningfully distinct from your existing {{Bad AI}} template? Belbury (talk) 19:39, 19 November 2023 (UTC)Reply[reply]

Wikimedia Foundation position on AI-generated content[edit]

The Wikimedia Foundation recently submitted some comments to the US Copyright Office in response to a Request for Comments on Artificial Intelligence and Copyright. Many of the points made by the Foundation will likely be of interest here, particularly the opening statement that:

Overall, the Foundation believes that generative AI tools offer benefits to help humans work more efficiently, but that there are risks of harms from abuse of these tools, particularly to generate large quantities of low-quality material.

File:Wikimedia Foundation’s Responses to the US Copyright Office Request for Comments on AI and Copyright, 2023.pdf

Omphalographer (talk) 20:06, 9 November 2023 (UTC)Reply[reply]

I wonder if there was a specific DR that made thw Foundation concerned about low quality spam. Or maybe someone just complained to staff staff? Trade (talk) 23:38, 18 November 2023 (UTC)Reply[reply]

Interesting[edit]

https://twitter.com/Kyatic/status/1725120435644239889 2804:14D:5C32:4673:DAF2:B1E3:1D20:8CB7 03:31, 17 November 2023 (UTC)Reply[reply]

This is highly relevant - thanks, whoever you are! Teaser:

This is the best example I've found yet of how derivative AI 'art' is. The person who generated the image on the left asked Midjourney to generate 'an average woman of Afghanistan'. It produced an almost carbon copy of the 1984 photo of Sharbat Gula, taken by Steve McCurry.

If you don't have a Twitter account, you can read the thread at https://nitter.net/Kyatic/status/1725120435644239889.
Omphalographer (talk) 03:59, 17 November 2023 (UTC)Reply[reply]
Hypothetically would we be allowed to upload the AI photo here? Trade (talk) 23:35, 18 November 2023 (UTC)Reply[reply]
No? The whole discussion is about how it's a derivative work of the National Geographic photo. Omphalographer (talk) 02:15, 19 November 2023 (UTC)Reply[reply]
I couldn't find clear info regarding artworks that look very close to non-CCBY photographs at Commons:Derivative works. This particular image may be fine, it's not prohibited just because the person looks similar to an actual person who was photographed and that photograph was the 'inspiration' to the AI creator.
That AI images look very similar to an existing photograph is overall exceptional and depends on issues with training data, parameters/weighting, and the prompts. Moreover, it's possible that this was caused on purpose to make a point or that they had put an extreme weight on high-valued photographs for cases like this while there are only few images of women from Afghanistan in the training data...more likely though the AI simply does not 'understand' (or misunderstands) what is meant by "average" here.
img2img issues
The bigger issue is that you can use images as input images and let AI modify them according to your prompt (example of how this can be really useful). This means some people may upload such an image without specifying the input image so people can check whether or not that is CCBY. If strength of the input image is configured to be e.g. 99% the resulting image would look very similar. I think there should be a policy that when you upload a AI-generated image via img2img, you should specify the input image. Prototyperspective (talk) 11:22, 19 November 2023 (UTC)Reply[reply]
If a human had created that image, we would certainly delete it as a plagiaristic copyvio. I see no reason to treat it more favorably because the plagiarist is a a computer program. - Jmabel ! talk 18:48, 19 November 2023 (UTC)Reply[reply]
I don't think so. I don't see much use of discussing this particular case and only meant to say that the derivative works does not really have info on this but I think artistic works that show something that has previously been photographed are allowed. Or are artworks of the Eiffel tower not allowed if the first depiction of it is a photograph that is not CCBY? Prototyperspective (talk) 18:58, 19 November 2023 (UTC)Reply[reply]
COM:BASEDONPHOTO is the relevant Commons policy for a drawing based on a single photograph: it requires the photographer's permission. I too would see no copyright difference between a human sketching a copy of a single specific photograph and an AI doing the same thing digitally. Belbury (talk) 19:19, 19 November 2023 (UTC)Reply[reply]
Thanks, a link to this policy was missing here so far. I don't see how the issue of photographing things like the Eiffel tower to copyright them away is addressed though. In this case, a person was photographed. I object to the notion that if the first photograph of a person, animal, object, or whatever is not in the public domain, it also can't be drawn under CCBY. I too would not see a copyright difference between a human sketching of a single photograph and an AI doing the same thing digitally. That is what img2img is, the case above is not based on a single image but many images, including many images of women. It would never work if it was just on one image. Prototyperspective (talk) 22:01, 19 November 2023 (UTC)Reply[reply]
The case above is not based on a single image but many images, including many images of women... I'm not convinced. The results shown look very much like they are based primarily on the National Geographic photo, possibly because there were many copies of it in the training data. Omphalographer (talk) 22:18, 19 November 2023 (UTC)Reply[reply]
@Prototyperspective: "the notion that if the first photograph of a person, animal, object, or whatever is not in the public domain, it also can't be drawn under CCBY." That's a straw-man argument, as is your Eiffel Tower example. You are refuting a claim that no one is making. While we have little insight into the "mind" of a generative AI system, I think we can reasonably conclude that if the AI had not seen that particular copyrighted image or works derived from it, then the chance is vanishingly small that it would have produced this particular image. And that is the essence of plagiarism. - Jmabel ! talk 01:06, 20 November 2023 (UTC)Reply[reply]
You may be misunderstanding COM:BASEDONPHOTO. It isn't saying that once somebody takes a photo of a subject, they have control over anyone who chooses to sketch the same subject independently in the future. It only applies to someone making a sketch using that photo alone as their reference for the subject.
The "many images" point doesn't seem very different from how a human would approach the same copying process. The human would also be applying various internal models of what women and hair and fabric generally look like, when deciding which details to include and omit, and which textures and styles to use. It would still result in a portrait that had been based closely on a specific existing one, so would be infringing on the original photographer's work. Belbury (talk) 13:01, 20 November 2023 (UTC)Reply[reply]
Now there are many good points here.
I won't address them in-depth or make any statements in regards to whether I agree with these points and their conclusion. Just a brief note on an issue & unclarity: what if that photo is the only photo of the organism or object? Let's say you want to draw an accurate artwork of an extinct animal photographed once where you'd orient by/use the photo – I don't think current copyright law finds you are not allowed to do so. In this case, I think this photo is the only known photo of this woman whose noncopyrighted genetics further emphasize her eyes making a certain noncopyrighted facial expression. Prototyperspective (talk) 13:18, 20 November 2023 (UTC)Reply[reply]
@Prototyperspective: I am going to assume good faith, and that you are not just arguing for the sake of arguing, but this is the last time I will respond here.
  • If there is exactly one photo (or other image) of a given organism or object, and it is copyrighted, and you create a work that is clearly derivative of it in a degree that is clearly plagiaristic, then most likely you are violating copyright. Consider the Mona Lisa. We don't have any other image of that woman. If it were a painting recent enough to still be in copyright, and you created an artwork that was nearly identical to the Mona Lisa, you'd be violating Leonardo's [hypothetical] copyright.
  • For your "extinct animal" case: probably the way to create another image that did not violate copyright would be to imagine it in a different pose (based at least loosely on images of a related species) and to draw or otherwise create an image of that. But if your drawing was very close to the only known image, and that image was copyrighted, you could well be violating copyright.
  • Again: the user didn't ask the AI to draw this particular woman. They asked for "an average woman of Afghanistan," and received a blatant plagiarism of a particular, iconic photo. Also, you say, "I think this photo is the only known photo of this woman." I suppose that may be an accurate statement of what you think, but it also tells me you have chosen to listen to your own thoughts rather than do any actual research. It is not the only photo of Sharbat Gula, nor even the only published photo of her. Other photos from that photo session when she was 12 years old were published (though they are less iconic) and I have seenn at least two published photos of her as an adult (one from the 2000s and one more recent). I suspect there are others that I have not seen. {[w|Sharbat Gula|She's had quite a life}} and now lives in Italy.
Jmabel ! talk 21:16, 20 November 2023 (UTC)Reply[reply]
This exact issue is described in Commons:AI-generated media#Copyrights of authors whose works were used to train the AI. It isn't discussed in other Commons policies because those documents were generally drawn up before AI image generation was a thing. Omphalographer (talk) 19:42, 19 November 2023 (UTC)Reply[reply]
Similarly to the issues presented in the Twitter/X thread, there is a lawsuit of a group of artists against several companies (incl. Midjourney and Stability AI), where a similar concern is presented (i.e. AI-generated media taking precedence above directly related images and works). I think this, among other things, is important to consider when deciding what scope Commons has in regards to AI-generated media. EdoAug (talk) 12:57, 9 December 2023 (UTC)Reply[reply]

Archiving[edit]

I would like to set up archiving fo this talk page to make active discussion more clearly visible. What do you think of a 8 month threshold of inactivity for now? By my estimation that would archive about a half or less of this talk page. Of course, the setting can be revisited later. whym (talk) 09:21, 25 November 2023 (UTC)Reply[reply]

i made one with 300d, which should let discussions stay around for about 1 year. :) --RZuo (talk) 14:35, 25 November 2023 (UTC)Reply[reply]
I think one year is by far not long enough to keep discussions. This should be much longer. It needs to be possible to figure out why a certain rule was implemented and what the substantiation was. JopkeB (talk) 04:41, 9 December 2023 (UTC)Reply[reply]
@JopkeB: It's not like you can't look at the archive. - Jmabel ! talk 08:20, 9 December 2023 (UTC)Reply[reply]
Thanks, User:Jmabel, I did not understand that "let discussions stay around for about 1 year" means that discussions stay on this talk page for about a year and then were moved to the archive. Of coarse this is OK. JopkeB (talk) 09:44, 9 December 2023 (UTC)Reply[reply]

Something I would actively like us to do involving AI images[edit]

I think it would be very useful to track the evolution of generative AI by giving some specific collection of prompts identically to various AI engines and see what each produces, and repeating this experiment periodically over the course of years to show how those engines evolve. Now that would clearly be within scope, if we don't trip across copyright issues of some sort. - Jmabel ! talk 21:55, 10 December 2023 (UTC)Reply[reply]

No need to highlight that being more clearly within scope here. I don't think you thought long about all the many potential and already existing use-cases of AI generated media especially once these improve. At some point with enough expertise you possibly could generate high-quality images of anything you can imagine matching quite closely to what you intended to depict (if you're skilled at prompting & modifying); how people don't see a tremendous potential in that is beyond me.
One part of that has already been done. However, that one implementation only compares one specific prompt at one point in time.
It would also be interesting to see other prompts such as landscape and people scenes prompts as well as how the generators improve on known issues such as misgenerated hands or basic conceptual errors (where e.g. people looking exactly the same are generated multiple times instead of only once as prompted). I think instead of uploading many images of the same prompt using different styles (example) it would be best to upload large collages (example) that include many styles/images at once. Since there is such a large number of styles, such applications are not as valuable as some other ones where one or a few images are enough and may even close a current gap. Prototyperspective (talk) 14:46, 13 December 2023 (UTC)Reply[reply]
For collages/montages we still prefer to also have each individual image available in a file of its own. - Jmabel ! talk 19:03, 13 December 2023 (UTC)Reply[reply]

More food for thought on AI-generated content[edit]

https://www.404media.co/facebook-is-being-overrun-with-stolen-ai-generated-images-that-people-think-are-real/

Omphalographer (talk) 18:02, 18 December 2023 (UTC)Reply[reply]

I think what they're posing there is more of a moral than a copyright question. What is being "stolen" is the concept of "man with wood-carved dog". Possibly none of the AI-generated images presented there is close enough to the original photos (also shown) to be considered a derivative work (I'm not quite sure, but the general characteristics of a bulldog or a German Shepherd aren't copyrighted, and the AI-works could be seen as just another interpretation of the concept - and mere concepts aren't copyrighted). But that doesn't mean that the moral question is uninteresting or unimportant. It's something we should consider when we talk about what we want to host on Commons. I previously advocated for keeping AI-generated content here only if it's in use in a Wikimedia project, but this makes me think: Imagine that Wikimedia projects start using AI-generated images of this kind to show something they otherwise couldn't show for copyright reasons? Thanks to fair use, some projects like English-language Wikipedia can show a lot more than others, for example, en:The Empire of Light does contain the works of Magritte under the fair use policy. But as German-language Wikipedia doesn't accept fair use (due to the absence of that legal provision in German-language countries), de:Das Reich der Lichter only links to images and doesn't show them directly. Well, now a German-language Wikipedian could tell the AI to generate a "painting of a house surrounded by trees at night, but under a daylight sky" and the result would probably be a "painting" that is similar to Magritte's works in its concept. It could then be used with a caption like "AI-generated approximation of Magritte's Empire of Light series", maybe without being a copyright violation - but I think we wouldn't want that? Gestumblindi (talk) 23:45, 18 December 2023 (UTC)Reply[reply]
The generated images in question here go beyond copying a "concept". They're being generated as fairly close replicas of specific, identifiable source photos; the article has examples.
Whether an "approximation" of a painting would be acceptable is questionable. If it's a close enough approximation, especially if it's imitating a specific work of art, that's likely to push it over the edge into being a derivative work. It's also questionable whether those projects would consider such an image educationally useful. Omphalographer (talk) 00:20, 19 December 2023 (UTC)Reply[reply]
I see now that some of the images are indeed very close to some of the originals (including virtually the same trees in the background, for example), but for others I would still say that they follow the concept without necessarily being a derivative work for copyright purposes. It would be legal (though not terribly original) to sculpt your own wooden dog and pose in a similar way as Michael Jones does. The very first image on the page, for example, has a background that is completely different from all of the photos by Jones shown, the dog is "sculpted" very differently, and the man looks nothing like Michael Jones. Gestumblindi (talk) 01:19, 19 December 2023 (UTC)Reply[reply]
"but I think we wouldn't want that?" It's not Commons responsibility to decide what images German Wikipedia users can use anymore than it's Commons responsibility to pick what admins they should have Trade (talk) 18:29, 20 December 2023 (UTC)Reply[reply]
I meant "we" more generally in the sense of the Wikimedia communities, not just Commons. I would strongly suspect that German-language Wikipedia's community would also be against such an approach. Gestumblindi (talk) 19:53, 20 December 2023 (UTC)Reply[reply]