Certainly there's an aspect of people using the chat interface like they use google: describe xyz to try to surface the name of a movie. Just in this case, we're doing the (less common?) query of: find me the picture I can vaguely describe; but it's a query to a image /generating/ service, not an image search service.
So I asked it to make 4 random and generic superheroes. It created Batman, Supergirl, Green Lantern, and Wonder Woman. Then at about 90% finished it deleted the image and said I was violating copyright.
I doubt the model you interact with actually knows why the babysitter model rejects images, but it claims to know why and leads to some funny responses. Here is it's response to me asking for a superhero with a dark bodysuit, a purple cape, a mouse logo on their chest, and a spooky mouse mask on their face.
> I couldn't generate the image you requested because the prompt involved content that may violate policy regarding realistic human-animal hybrid masks in a serious context.
(hard to formulate why I was too lazy to test myself :) )
I have a strong suspicion that many human artists would behave in a way the chat bot did (unless they start asking clarifying questions. Which chatbots should learn to do as well)
So the criminal party here would be OpenAI, since they are selling access to a service that generates copyright-infringing images.
Ironically that's probably because the errors and flaws in those generations at least made them different from what they were attempting to rip off.
Overall the model is tra
I wonder if it's a fine tuning issue where people have overly provided archetypes of the thing that they were training towards. That would be the fastest way for the model to learn the idea but it may also mean the model has implicitly learned to provide not just an instance of a thing but a known archetype of a thing. I'm guessing in most RLHF tests archetypes (regardless of IP status) score quite highly.
ClosedAI doesn't seem to be OK with it, because they are explicitly censoring characters of more popular IPs. Presumably as a fig leaf against accusations of theft.
The question is "should we define it as such?"
But because a computer, and not a human does it, they get to launder their responsibility.
For humans it doesn't make sense because we have generation and filtering in a single package.
Overfitting is if you didn't exactly describe Indiana Jones and then it still gave Indiana Jones.
It didn't though, it just spat out what is basically a 1:1 copy of some Indiana Jones promo shoot. No where did the prompt ask for it to look like Harrison Ford.
If we were playing Charades, just about anyone would have guessed you were describing Indiana Jones.
If you gave a street artist the same prompt, you'd probably get something similar unless you specified something like "... but something different than Indiana Jones".
How can you express, in term of AI training, ignoring the existence of something that's widely present in your training data set? if you ask the same question to a 18yo girl in rural Thailand, would she draw Harrison Ford as Indiana Jones? Maybe not. Or maybe she would.
But IMO an AI model must be able to provide a more generic (unbiased?) answer when the prompt wasn't specific enough.
Maybe it would have some point if you are targetting users in a substantially different social context. In the case, you would design the model to be familiar with their tropes instead. So when they describe a character iconic in their culture, by a few distinguishing characteristics, it would produce that character for them. That's no different at all.
But if you look at it from the perspective that there is only one example to learn, from it is maybe not over it.
That's not overfitting. That's either just correct or underfitting (if we say it's never returning anything but 2)!
Overfitting is where the model matches the training data too closely and has inferred a complex relationship using too many variables where there is really just noise.
Recent benchmark on unseen 2025 Math Olympiad shows none of the models can problem solve . They all accidentally or on purpose had prior solutions in the training set.
[0] https://imgur.com/a/wqrBGRF Image captions are the impled IP, I copied the prompts from the blog post.
To me, this article is further proof that LLMs are a form of lossy storage. People attribute special quality to the loss (the image isn't wrong, it's just got different "features" that got inserted) but at this point there's not a lot distinguishing a seed+prompt file+model from a lossy archive of media, be it text or images, and in the future likely video as well.
The craziest thing is that AI seems to have gathered some kind of special status that earlier forms of digital reproduction didn't have (even though those 64kbps MP3s from napster were far from perfect reproductions), probably because now it's done by large corporations rather than individuals.
If we're accepting AI-washing of copyright, we might as well accept pirated movies, as those are re-encoded from original high-resolution originals as well.
To a viewer, a human-made work and an AI-generated one both amount to a series of stimuli that someone else made and you have no control over; and when people pay to see a movie, generally they don't do it with the intent to finance the movie company to make more movies -- they do it because they're offered the option to spend a couple hours watching something enjoyable. Who cares where it comes from -- if it reached us, it must be good, right?
The "special status" you speak of is due to AI's constrained ability to recombine familiar elements in novel ways. 64k MP3 artifacts aren't interesting to listen to; while a high-novelty experience such as learning a new culture or a new discipline isn't accessible (and also comes with expectations that passive consumption doesn't have.)
Either way, I wish the world gave people more interesting things to do with their brains than make a money, watch a movies, or some mix of the two with more steps. (But there isn't much of that left -- hence the concept of a "personal life" as reduced to breaking one's own and others' cognitive functioning then spending lifetimes routing around the damage. Positively fascinating /s)
A new MCU movie is released, its 60 second trailer posted on Youtube, but I don't feel like watching the movie because I got bored after Endgame.
Youtube has very strict anti-scraping techniques now, so I use deep-scrapper to generate the whole trailer from the thumbnail and title.
I use deep-pirate to generate the whole 3 hour movie from the trailer.
I use deep-watcher to summarize the whole movie in a 60 second video.
I watch the video. It doesn't make any sense. I check the Youtube trailer. It's the same video.
I feel like the less advanced generations, maybe even because of their limitations in terms of size, were better at coming up with something that at least feels new.
In the end, other than for copyright-washing, why wouldn't I just use the original movie still/photo in the first place?