Discussion:Soucougnan

Art créé par intelligence artificielle

@Triton Firstly, my apologies for not writing in French on the French Wikipedia. Je ne parle pas français. Hopefully you are able to put my comment through something like Google Translate.

I do not agree with including AI-generated art on Wikipedia articles, as doing so is against Wikipedia's verifiability policy (vérifiabilité). The result generated from the prompt "a witch looking like a big black bird under a kapok tree in Guadeloupe" is not enough to properly illustrate this subject, as the AI engine does not need to generate something that actually resembles a soucouyant (soucougnan).

Why is she wearing a witch's hat, something that doesn't exist in Caribbean folklore? What is the blue pendant supposed to be? Why are crows relevant? What is the white stuff around her head? All these questions are pointless to ask, because artificial intelligence cannot tell you how it thinks, and is not verifiable.

The article is better off having no illustration at all until an actual human artist can create an image that is actually based on the features of this creature, rather than an AI which is prone to hallucinating anything and can generate all manners of false results. Opencooper (discuter) 8 novembre 2023 à 09:02 (CET)

Hi! No problem with the langage. I'm sorry but there is absolutely to guidelines about IA illustration on such subject, and the general use of AI for this has already been discussed with no real opposition. While there can be discussion picture by picture if there is questions about some features, I can't accept a deletion based solely on the fact that it's an AI illustration. Please take into consideration that I don't need consensus to add a picture on an article like implied in the comments (like everyone, I'm entitled to add new stuff if it respects the rules) but you need consensus before deleting several times the modification of another user.

So, back to this specific picture: to be honest, I wasn't totally happy with each details of this one and can work to improve it. The hat may not be appropriate indeed. The link with the crows are just a reminder of the link between this creature and birds, the blue pendant is totally anecdotal and the white stuff simply braids. So, as I said, I will work toward a better illustration, been trough AI ou trough manual adjustement, as always. Please do not rv again unless you have substantial element that oppose this specific picture, more than just the hat. Triton (discuter) 8 novembre 2023 à 23:07 (CET)

@Triton My opposition isn't solely because it is AI-generated, but because AI-generated images lack verifiability, and in this specific case, it has been demonstrated that the AI completely made up the elements of the picture under discussion. If you search Google Images for "soucouyant", not a single result will show a witch's hat. My reason for pointing this out isn't because the hat is my only problem with the image, but because it points towards the wider issue that: AI is known to generate false information, and thus there is no guarantee that this image accurately illustrates a soucouyant. No human illustrator would draw that hat, because it's not an element of Caribbean folklore. That's just one obvious place where the AI is wrong, but one could nitpick the image all day.

Since you want more than just the hat, there are many more problems with the depiction. Why is her face in shadow during the daytime? Why are her eyes lit up? She's supposed to look like an old lady in her normal form. You say the white things are braids, but they look like bandages to me. (braids aren't arrow/fish shaped) Her thumb appears broken, and her fingers are all the same length. Saying you'll just generate another image ignores the limitations and reality of this technology.

The image is not created on a basis of what a soucouyant is. It's just whatever random elements the AI chose to output. I'm really surprised that you don't see a problem with including something like this on an encyclopedia project that has strict guidelines for variability of information, especially after problems have already been pointed out with it. Opencooper (discuter) 9 novembre 2023 à 00:17 (CET)

I'm sorry but your arguments are mostly aesthetic. An AI picture is indeed no perfect but we are not speaking about perfection here, we are speaking about an "artistic" depiction of folkloric being for which there is no exact description (or rather several descriptions). You can basicaly have the same position for about any illustration on Wikipédia which is not a photograph. Of course some old painting may have some notoriety that may have a self-explanatory title, but it's still an interpretation of the legend. If we had and illustration posted somewhere else under free licence and clearly title "soucougnan", would it be better? And if we had an illustration of a soucougnan from a children book (for example, made by miracle into open licence) made by a professionnal illustration would it be better? There may be some better attention to details but it would be exactly the same: it would never accurately illustrate an soucougnan since it doesn't exist.

Side note: French WP has made a community discussion about using drawings for portraits, which is specific case for which I agree AI could not (at this time of the technology) be used. But for this kind of topic, the objective of an illustration is to show an interpretation of the subject and AI can be a tool if correctly used. I think you imagine that I just typed the prompt and took the first image, but it's not the case, it took a lot of iterations to get a correct image, with some automatic and some manual correction.

But, as I said, I wasn't totally happy with the final result, and considering your comments I started again some iteration and manual correction to have a new picture you can now see on the article. I hope it will suit you better. If you have more problem with it I invite you to launch a larger discussion onthe use of AI for fictionnal characters, since the points you consider problematics are not specific to this article. Triton (discuter) 11 novembre 2023 à 17:06 (CET)

@Triton Thank you for the response. In response to your questions about a possible freely illustrated depiction from somewhere else or a children's book being preferable, my answer would be a resounding "yes". The difference is that a human is using their knowledge and judgement to portray the subject, rather than an AI making things up without a basis. Prompting is not the same thing as thinking. AI image generation is just sophisticated pattern matching, hence it does not actually understand for example that a human hand has five fingers, but rather the random noise it uses for generation sometimes coincidentally ends up with a full set of fingers, and many more times less. The subject being depicted being fictional does not nullify these concerns.

It's pointless arguing over the merits or demerits of a new image, because my stance was always that usage of an AI generated image isn't acceptable for me (with no illustration being preferable to having an AI one). I would start a wider discussion, but my unfamiliarity with the language and norms of the French Wikipedia make me hesitant to do so. On the English Wikipedia at least, I would know WikiProjects or specific talk pages to point towards this discussion to get more participants, but I'm out of my depth here, so I will refrain. If you think having more viewpoints here would enrich the discussion and know the forums to solicit this, I invite you to please do so. Opencooper (discuter) 11 novembre 2023 à 22:27 (CET)

So your opininon is really just about AI in principle, not about the quality of the result. As I said, you have no idea the steps I took to create the picture, and not only it did come from thinking (with a lot -really- of combinations of words and choice of patterns from me), same as a man-made illustration, but it also was the result from AI-generated details generation (you can chose some specific areas to try new combinations) but also manual editing on my part in several iteration and in the end. So no, it is not an IA interpretation (as you wrote) but an IA-helped illustration (you can add Gimp-helped if you want to be complete at this point) interpretation from me. I invite you to test how to get to such result before judging the tool(s). Otherwise I can't see how we can constructively argue.

I'm sorry but I don't have the time to launch and manage a discussion about this topic with the whole community. I already got postive feedbacks on other contribution of the sort and no critics before you. If this really concerns you, you can make a short message in english on the village pump, redirecting to this conversation, there will be enough english-speaking contributors to understand if they want to continue. I will be happy to answer anyone here in french who comes with questions. Triton (discuter) 12 novembre 2023 à 01:14 (CET)