Will a significant AI generated meme occur before 2025?
➕
Plus
51
Ṁ4586
Mar 1
44%
chance

This question resolves as YES if, before the January 1st 2025, a meme generated by artificial intelligence gains significant recognition and viral status, comparable to the widespread influence and shareability of popular memes created by humans. The meme must be widely acknowledged as being AI-generated and should achieve a level of cultural impact and dissemination across various media and social platforms, reflecting a substantial influence in the meme culture and digital community.

Questions with the same criteria:

/RemNi/will-a-significant-ai-generated-mem-5f798d3b2ace (this question)

/RemNi/will-a-significant-ai-generated-mem-1760ddcaf500

/RemNi/will-a-significant-ai-generated-mem-200a1b818595

/RemNi/will-a-significant-ai-generated-mem-95b65e96092d

/RemNi/will-a-significant-ai-generated-mem-7b1d4b5ec34

/RemNi/will-a-significant-ai-generated-mem-35fc023d619f

/RemNi/will-a-significant-ai-generated-mem

Other questions for 2025:

/RemNi/will-we-get-agi-before-2025

/RemNi/will-we-get-room-temperature-superc-e940f30870be

/RemNi/will-we-discover-alien-life-before-031ec0858fcc

/RemNi/will-a-human-walk-on-the-moon-again-97ccfb75fd6d

/RemNi/will-we-get-fusion-reactors-before-d18e9fd38cd1

/RemNi/will-we-get-a-cure-for-cancer-befor-bf2acb801224

The "idea" of the meme must have been discovered by an AI. If a human comes up with an idea for a meme, and then creates an image using a text-to-image model and puts it online, this does not count. Furthermore, the meme cannot simply be the fact that the content itself is AI generated. The semantics of the meme must refer to something else. There must be strong evidence that the AI chose the idea for the meme for this question to resolve as YES. If there is weak evidence, but no strong evidence, then this question resolves as N/A. If there is no evidence that this has occurred by the end date, then this question resolves as NO.


The meme must not solely be text-based, and must contain a non-text AI-generated component (e.g. image, video or audio). Text-based AI-generated memes arguably already exist:

https://knowyourmeme.com/memes/ai-generated-greentexts

Note that the meme must become significant before this question's end date. If an AI-generated meme is created in a given year, but only becomes popular in the following year, then it is the following year's question that resolves as YES.

Get
Ṁ1,000
and
S3.00
Sort by:

@traders The wojack meme generator that's making the rounds on twitter (from https://glif.app/glifs) only counts if the topic ("idea") was chosen by the AI. If a human writes a prompt for a specific type of person, and then the resulting meme becomes popular, then this does not count.

some of them are pretty good and it's looking like we're getting closer to a YES resolution of this question given these capabilities (think this is sonnet 3.5?)

Some candidates here https://old.reddit.com/r/midjourney/comments/1brf076/so_i_asked_ai_to_generate_me_some_memes_uhh/ (although its unclear to what extent OP specifically prompted for those designs)

@retr0id Any evidence that these spread beyond that reddit post?

@RemNi What do you think of The Unknown, a completely AI-generated character from the recent Willy Wonka Experience?

bought Ṁ50 YES from 51% to 58%

@wasabipesto Uh huh. An "actor playing the part of an AI-generated character at a willy-wonka themed scam event in Scotland" wasn't on my radar as being the first "significant AI generated meme" when I wrote this question. I'm going to have to think about this one for a bit.

@wasabipesto My thoughts so far:

- From what I've seen online it does appear to meet the criteria of "the idea of the meme must have been discovered by an AI", since it looks like the AI has hallucinated a character that doesn't have a whole lot to do with willy wonka.

- Most of its popularity online (as of March 3rd) seems to be coming from its weirdness independent of people being aware that it's AI generated. As indicated in these tweets: https://twitter.com/bene25_/status/1762631362597519859 , https://twitter.com/dvd_brwn/status/1762809446751687105 , https://twitter.com/JakeMC_/status/1762898809007313310

- The meme is not solely text-based in that it involves a part played on a "stage". However this stage performance is not AI-generated. It's possible that the AI-generated bit here is just the text component, in which case this would not qualify since it doesn't exceed AI-generated greentexts.

I would, however, be curious as to how the organisers came up with the costume idea for The Unknown. If they put the description through a text-to-image generator, which then produced an image of the costume that they then went about imitating, this meme would qualify with respect to this question's criteria. Because the AI-generated meme would then be "the generated text with the hallucinated character" + "the generated image with the character's costume". The stage performance would in this (hypothetical scenario) be downstream of this AI-generated meme.

@wasabipesto I won't resolve this question as YES without stronger evidence. It definitely IS possible that this meme qualifies, but there's currently a lack of evidence supporting this. Will be cool if someone does a bit more digging / forensics to figure out where this came from, particularly the costume design (maybe there's some midjourney image somewhere with corresponding keywords)

In any case really interesting example!

@RemNi My understanding is that the stage directions and costumes were part of the (AI-generated) script, with one exception - the actor improvised hiding behind the mirror according to this interview. I think you could easily break it down to:
A: The meme is the character, whose identity and appearance were (as I understand) both AI-generated but whose defining behavior was not. The character design itself is moderately popular, but likely as a result of the event being popular and not of its own right. I think this meets ~70% of the criteria.
B: The meme is the performance, which was inspired by AI but performed by an actor. I think this does not really meet any of the criteria.

Thanks for looking into it!

@wasabipesto There's a hint that they used midjourney: https://www.midjourney.com/jobs/726699db-3584-4e69-89b8-f2931352d40e?index=0

(from 4 weeks ago)

Although this could just be a coincidence and someone else was doing a candy themed amusement park around the same time.

@wasabipesto I saw this image on twitter:

https://twitter.com/AlsikkanTV/status/1762235022851948668

But not sure they used midjourney to generate that.

@wasabipesto https://twitter.com/dswhite7/status/1762280216670707715

This looks like they were using dalle-3 and not midjourney

Can you elaborate on the restrictions of how the meme must be generated? For example bottomless pit supervisor starts with a prompt, so would that fail because a human came up with the idea?

What about dripped up pope, or will smith eating spaghetti? Neither example show evidence of their OPs creating them with any intention to become a meme, but both were spread, gained fame and remixed.

@qqii For bottomless pit supervisor, my understanding is that the prompt was limited to something along the style of "generate a 4chan greentext", in which case the resulting output qualifies since it became a (fairly) widespread meme. For a meme to resolve this question, a non-text (e.g. image, video or audio) component is required.

For dripped up pope and will smith eating spaghetti, the issue is not the OP intent with respect to the final popularity, but that they came up with the idea of the meme. If there was strong evidence that the user had just put in "generate a meme with an image" to gpt4, and it came up with "pope wearing balenciaga" and generated the corresponding image, then it would qualify. In the absence of evidence that this took place, these examples do not qualify.

(also strongly assume that for dripped up pope and will smith eating spaghetti there was significant human authorship and iteration on the idea before the final version that we saw).

@RemNi thanks for carification. Bottomless pit supervisor's prompt is clearly shown in the original image since it is devoid of green highlighting:

"""

write me a 4chan greentext

>be me

>bottomless pit supervisor

>

"""

So I'm guessing it doesn't count because the idea of "bottomless pit supervisor" was supplied by a human?

I'm still not clear how you distinguish "idea of the meme" versus just creating something. Not all ideas become memes, and my point is that the author of at least will smith eating spaghetti had no intention for it to be remixed - which I think is a more important factor to distinguish a meme than just popularity alone.

I struggle to see evidence of iteration either, as the output is no more than what modelscope text2video could achieve - evidenced by the multitude of variants people generated. Unlike pope drip, would you not consider the viral element of it a complete fabrication of (the limitations of) AI? Clearly the author did not intend for what created when they gave the prompt "Will Smith eating spaghetti.", and it was only through (the deficiencies of) AI that such a normally simple and mundane statement could become the cultural phenomena it is now where even the actual will smith eating spaghetti has meaning.

@qqii good point about the green highlighting. On the original one, the green highlighting is absent both at the start and the end of the text, so not 100% sure that this highlighting corresponds to the "AI-generated" part.

There are a few AI-generated greentext memes here: https://knowyourmeme.com/memes/ai-generated-greentexts . Admittedly it is hard to know for sure which ones contain ideas "created" by the AI and which ones are prompted by the user, since the highlighting doesn't necessarily reflect the "AI-generated" part. I think 2 years ago on the OpenAI playground the default behavior for gpt-3 was to add this style of highlighting: https://www.technologyreview.com/2022/08/31/1058800/what-does-gpt-3-know-about-me/

But yes, I might edit and clarify the description of this question a bit because the bottomless pit supervisor example is confusing if it is the case that the idea was provided by the human as the highlighting suggests. Thanks for the tip.

As to your second question about the "idea of the meme" vs "just creating something", I think there is a misunderstanding somewhere. In the case of will smith eating spaghetti, I'm 95% sure that a human had that idea one day in early 2023, then sat at their computer for >4 hours to turn that into reality using various text-to-image/text-to-video tools. In this case, there is no strong evidence that the AI came up with the "idea of the meme". It's very likely that the human put into the text-to-image or text-to-video generator the prompt "will smith eating spaghetti", and then selected the best samples and concatenated them together into the final video.

Hypothetically, if a human had just put into their AI-agent (something like autogen) a prompt along the lines of "come up with a video meme", and let it spin overnight. If that process had produced the original will smith eating spaghetti and we had strong evidence that this occurred, then that video would qualify. Since our confidence is high that this did not take place, and a human came up with the idea and not the AI, therefore this video does not qualify.

Does that make sense?

@RemNi yeah, it's highlighting from OpenAI playground. You can make edits any any point and ask it to continue generating - and you can see that in MIT article unhighlighted questions interspersed with highlighted answers.

I'm 95% sure that a human had that idea one day in early 2023, then sat at their computer for >4 hours to turn that into reality using various text-to-image/text-to-video tools. In this case, there is no strong evidence that the AI came up with the "idea of the meme". It's very likely that the human put into the text-to-image or text-to-video generator the prompt "will smith eating spaghetti", and then selected the best samples and concatenated them together into the final video.

You can try modelscope text to video yourself. From my experience, there was no ">4 hours" and whatever selection of samples was minimal at best, for I cannot distinguish from what I've randomly generated to the original. Each clip takes mere seconds to generate, and even if the best of double were selected that's still only a handful of minutes.

@qqii Ok sure, maybe minutes and not 4 hours. I believe I have answered your question, if you have any others please let me know.

@qqii And thanks for the tip about the greentexts!

@RemNi sorry, I don't think I'll be betting here as I still don't understand where you're drawing the line.

To me, it is clear that no human would seek out to create the idea of "will smith eating spaghetti" as its meme form. The meme is not literally "will smith eating spaghetti". This is clear by will smith's own parody. The idea of the meme is something more, and to my mind was clearly "discovered" rather than "invented".

Then in the other thread where you consider the unknown as discovered by AI is equally as confusing - for the core of the meme meme lies not with the script but with the costume, the performance, the reaction and the situation. The evidence is that Billy Coull set out to run an event, and organised an event with the assistance of AI. Clearly more effort, assistance and curation was given than /u/chaindrop.

@qqii fine by me if you don't bet on this market. I think you might be having trouble understanding the question description and my answers, which is also fine. I'm not 100% sure what you're getting at or the specific point you're trying to make upon re-reading your comments, but it sounds like you might be referring to memes as abstract concepts, as opposed to what the criteria of this question describes. You also appear to have different beliefs about how some existing memes were generated. I encourage you to write your own question with a different set of criteria that make sense from your perspective if this is the case. Not much value continuing this thread because at this point it's going in circles, but thanks for your input anyway!

I asked Bing Chat “Generate a random prompt for an image that could go viral, then generate the image”

Bing Chat came up with “A cat riding sunglasses and a hat riding a skateboard”, then generated these images:

So if any of image goes viral as a meme, this market resolves YES.

An AI generated meme might have already gone viral already, but it’s hard if not impossible to tell if the prompting was done by a human or an AI.

@soweliSon haha I'm all for it. Under the constraint that it has to become a significant meme, but I'm willing to believe.

And yes, it's possible that an AI generated meme may have gone viral already. This market leaves 3 months after Jan 1st 2025 for evidence to emerge that a significant AI generated meme occurred before 2025.

@RemNi Does the AI image of the Pope in a puffer jacket count?

predictedYES

@NathanShowell I believe it does not count because a human, presumably poster of this Reddit post the came up with the idea and the prompt, and not an AI.

https://www.reddit.com/r/midjourney/s/nHypXkqnne

@soweliSon yeah that's correct

@soweliSon the prompt is on this page: https://knowyourmeme.com/memes/pope-in-white-puffer-jacket-pope-francis-drip

But I think it's a good example to show that it seems like this question may resolve as YES in the not too distant future.

However, a lot of the interest in the meme at the time was that it was made with an image generator. Now that people have become more used to this notion, it's going to be more challenging for synthetic memes to become popular.

© Manifold Markets, Inc.Terms + Mana-only TermsPrivacyRules