This line of thinking is a real misunderstanding of creativity. “Humans are only remixing things they’ve learned, so AI is the same”. It’s not the same. If an AI has nothing but images of apples in its training data, it will never ever be able to draw a banana. It seems creative and smart on the surface only because its trained on the (stolen) input of every bit of art, text, and code on the internet. But if humans stop creating new imaginative input, it will stagnate right where it is.
If someone has never seen a banana they wouldn’t be able to draw it either.
Also, AIs aren’t stealing anything. When you steal something you have deprived the original owner of that thing. If anything, AIs are copying things but even that isn’t accurate.
When an image AI is trained, it reads though millions upon millions of images that live on the public Internet and for any given image it will increment a floating point value by like 0.01. That’s it. That’s all they do.
For some reason people have this idea in their heads that every AI-generated image can be traced back to some specific image that it somehow copied exactly then modified slightly and combined together for a final output. That’s not how the tech works at all.
Hard disagree. You just have to describe the shape and colors of the banana and maybe give it some dimensions. Here’s an example:
A hyper-realistic studio photograph of a single, elongated organic object resting on a wooden surface. The object is curved into a gentle crescent arc and features a smooth, waxy, vibrant yellow skin. It has distinct longitudinal ridges running its length, giving it a soft-edged pentagonal cross-section. The bottom end tapers toa small, dark, organic nub, while the top end extends into a thick, fibrous, greenish-brown stalk that appears to have been cut froma larger cluster. The yellow surface has minute brown speckles indicating ripeness.
It’s a lot of description but you’ve got 4096 tokens to play with so why not?
Remember: AI is just a method for giving instructions to a computer. If you give it enough details, it can do the thing at least some of the time (also remember that at the heart of every gen AI model is a RNG).
Note: That was the first try and I didn’t even use the word “banana”.
Totally disagree. I’ve seen original sources reproduced that show exactly what an AI copied to make images.
And humans can definitely create things that have never been seen before. An AI could never have invented general relativity having only been trained on Newtonian physics. But Einstein did.
I’ve seen original sources reproduced that show exactly what an AI copied to make images.
Show me. I’d honestly like to see it because it means that something very, very strange is taking place within the model that could be a vulnerability (I work insecurity).
The closest thing to that I’ve seen is false watermarks: If the model was trained on a lot of similar images with watermarks (e.g. all images of a particular kind of fungus might have come from a handful of images that were all watermarked), the output will often have a nonsense watermark that sort of resembles the original one. This usually only happens with super specific things like when you put the latin name of a plant or tree in your prompt.
Another thing that can commonly happen is hallucinated signatures: On any given image that’s supposed to look like a painting/drawing, image models will sometimes put a signature-looking thing in the lower right corner (because that’s where most artist signatures are placed).
The reason why this happens isn’t because the image was directly copied from someone’s work, it’s because there’s a statistical chance that the model (when trained) associated the keywords in your prompt with some images that had such signatures. The training of models is getting better at preventing this from happening though, as they apply better bounding box filtering to the images as a pretraining step. E.g. a public domain Audibon drawing of a pelican would only use the bird itself and not the entire image (which would include the artist signature somewhere).
The reason why the signature should not be included is because the resulting image would not be drawn by that artist. That would be tantamount to fraud (bad). Instead, what image models do (except OpenAI with ChatGPT/DALL-E) is tell the public exactly what their images were trained on. For example, they’ll usually disclose that they used ImageNET (which you yourself can download here: https://www.image-net.org/download.php ).
Note: I’m pretty sure the full ImageNET database is also on Huggingface somewhere if you don’t want to create an account with them.
Also note: ImageNET doesn’t actually contain images! It’s just a database of image metadata that includes bounding boxes. Volunteers—for over a decade—spent a lot of time drawing bounding boxes with labels/descriptions on public images that are available for anyone to download for free (with open licenses!). This means that if you want to train a model with ImageNET, you have to walk the database and download all the image URLs it contains.
If anything was “stolen”, it was the time of those volunteers that created the classification system/DB in order for things like OpenCV to work so that your doorbell/security camera can tell the difference between a human and a cat.
What that Afghanistan girl image demonstrates is simply a lack of diversity in Midjourney’s training data. They probably only had a single image categorized as “Afghanistan girl”. So the prompt ended up with an extreme bias towards that particular set of training values.
Having said that, Midjourney’s model is entirely proprietary so I don’t know if it works the same way as other image models.
It’s all about statistics. For example, there were so many quotes and literal copies of the first Harry Potter book in OpenAI’s training set that you could get ChatGPT to spit out something like 70% of the book with a lot of very, very specific prompts.
At the heart of every AI is a random number generator. If you ask it to generate an image of an Afghan girl—and it was only ever trained on a single image—it’s going to output something similar to that one image every single time.
On the other hand, if it had thousands of images of Afghan girls you’d get more varied and original results.
For reference, finding flaws in training data like that “Afghanistan girl” is one of the methods security researchers use to break large language models.
Flaws like this are easy to fix once they’re found. So it’s likely that over time, image models will improve and we’ll see fewer issues like this.
The “creativity” isn’t in the AI model itself, it’s in its use.
I guess the argument is “if the AI mixes enough copied art together so that you can’t tell as easily, it’s being creative like a human” and I just don’t really believe that. Perhaps its a philosophical question.
That vase was made using code (literally just text) I wrote in OpenSCAD. The model file is the result of the code I wrote and the physical object is the output of the 3D printer that I built. The pretty filament was store-bought, however.
If giving a machine instructions doesn’t count as a creative process then programming doesn’t count either. Because that’s all you’re doing when you feed a prompt to an AI: Giving it instructions. It’s just the latest tech for giving instructions to machines.
This line of thinking is a real misunderstanding of creativity. “Humans are only remixing things they’ve learned, so AI is the same”. It’s not the same. If an AI has nothing but images of apples in its training data, it will never ever be able to draw a banana. It seems creative and smart on the surface only because its trained on the (stolen) input of every bit of art, text, and code on the internet. But if humans stop creating new imaginative input, it will stagnate right where it is.
If someone has never seen a banana they wouldn’t be able to draw it either.
Also, AIs aren’t stealing anything. When you steal something you have deprived the original owner of that thing. If anything, AIs are copying things but even that isn’t accurate.
When an image AI is trained, it reads though millions upon millions of images that live on the public Internet and for any given image it will increment a floating point value by like 0.01. That’s it. That’s all they do.
For some reason people have this idea in their heads that every AI-generated image can be traced back to some specific image that it somehow copied exactly then modified slightly and combined together for a final output. That’s not how the tech works at all.
You can steal a car. You can’t steal an image.
“If someone hasn’t seen a banana they wouldn’t be able to draw it either.”
Hard disagree. If you gave said person a prompt about what a banana looks like they could draw one.
No matter how much prompting you do, an ai that hasn’t seen a banana cannot make one.
Hard disagree. You just have to describe the shape and colors of the banana and maybe give it some dimensions. Here’s an example:
A hyper-realistic studio photograph of a single, elongated organic object resting on a wooden surface. The object is curved into a gentle crescent arc and features a smooth, waxy, vibrant yellow skin. It has distinct longitudinal ridges running its length, giving it a soft-edged pentagonal cross-section. The bottom end tapers to a small, dark, organic nub, while the top end extends into a thick, fibrous, greenish-brown stalk that appears to have been cut from a larger cluster. The yellow surface has minute brown speckles indicating ripeness.It’s a lot of description but you’ve got 4096 tokens to play with so why not?
Remember: AI is just a method for giving instructions to a computer. If you give it enough details, it can do the thing at least some of the time (also remember that at the heart of every gen AI model is a RNG).
Note: That was the first try and I didn’t even use the word “banana”.
That’s also not a banana. It’s some type of gourde.
It’s close enough. The key is that it’s not something that was just regurgitated based on a single keyword. It’s unique.
I could’ve generated hundreds and I bet a few would look a lot more like a banana.
But also how many bananas has that llm you’re using seen? You’re sure it’s zero?
Totally disagree. I’ve seen original sources reproduced that show exactly what an AI copied to make images.
And humans can definitely create things that have never been seen before. An AI could never have invented general relativity having only been trained on Newtonian physics. But Einstein did.
Show me. I’d honestly like to see it because it means that something very, very strange is taking place within the model that could be a vulnerability (I work insecurity).
The closest thing to that I’ve seen is false watermarks: If the model was trained on a lot of similar images with watermarks (e.g. all images of a particular kind of fungus might have come from a handful of images that were all watermarked), the output will often have a nonsense watermark that sort of resembles the original one. This usually only happens with super specific things like when you put the latin name of a plant or tree in your prompt.
Another thing that can commonly happen is hallucinated signatures: On any given image that’s supposed to look like a painting/drawing, image models will sometimes put a signature-looking thing in the lower right corner (because that’s where most artist signatures are placed).
The reason why this happens isn’t because the image was directly copied from someone’s work, it’s because there’s a statistical chance that the model (when trained) associated the keywords in your prompt with some images that had such signatures. The training of models is getting better at preventing this from happening though, as they apply better bounding box filtering to the images as a pretraining step. E.g. a public domain Audibon drawing of a pelican would only use the bird itself and not the entire image (which would include the artist signature somewhere).
The reason why the signature should not be included is because the resulting image would not be drawn by that artist. That would be tantamount to fraud (bad). Instead, what image models do (except OpenAI with ChatGPT/DALL-E) is tell the public exactly what their images were trained on. For example, they’ll usually disclose that they used ImageNET (which you yourself can download here: https://www.image-net.org/download.php ).
Note: I’m pretty sure the full ImageNET database is also on Huggingface somewhere if you don’t want to create an account with them.
Also note: ImageNET doesn’t actually contain images! It’s just a database of image metadata that includes bounding boxes. Volunteers—for over a decade—spent a lot of time drawing bounding boxes with labels/descriptions on public images that are available for anyone to download for free (with open licenses!). This means that if you want to train a model with ImageNET, you have to walk the database and download all the image URLs it contains.
If anything was “stolen”, it was the time of those volunteers that created the classification system/DB in order for things like OpenCV to work so that your doorbell/security camera can tell the difference between a human and a cat.
Examples like this (first one I found just now):
More reading
https://www.kortizblog.com/blog/why-ai-models-are-not-inspired-like-humans
What that Afghanistan girl image demonstrates is simply a lack of diversity in Midjourney’s training data. They probably only had a single image categorized as “Afghanistan girl”. So the prompt ended up with an extreme bias towards that particular set of training values.
Having said that, Midjourney’s model is entirely proprietary so I don’t know if it works the same way as other image models.
It’s all about statistics. For example, there were so many quotes and literal copies of the first Harry Potter book in OpenAI’s training set that you could get ChatGPT to spit out something like 70% of the book with a lot of very, very specific prompts.
At the heart of every AI is a random number generator. If you ask it to generate an image of an Afghan girl—and it was only ever trained on a single image—it’s going to output something similar to that one image every single time.
On the other hand, if it had thousands of images of Afghan girls you’d get more varied and original results.
For reference, finding flaws in training data like that “Afghanistan girl” is one of the methods security researchers use to break large language models.
Flaws like this are easy to fix once they’re found. So it’s likely that over time, image models will improve and we’ll see fewer issues like this.
The “creativity” isn’t in the AI model itself, it’s in its use.
I guess the argument is “if the AI mixes enough copied art together so that you can’t tell as easily, it’s being creative like a human” and I just don’t really believe that. Perhaps its a philosophical question.
It’s more like this: If you give a machine instructions to construct or do something, is the end result a creative work?
If I design a vase (using nothing but code) that’s meant to be 3D printed, does that count as a creative work?
https://imgur.com/bdxnr27
That vase was made using code (literally just text) I wrote in OpenSCAD. The model file is the result of the code I wrote and the physical object is the output of the 3D printer that I built. The pretty filament was store-bought, however.
If giving a machine instructions doesn’t count as a creative process then programming doesn’t count either. Because that’s all you’re doing when you feed a prompt to an AI: Giving it instructions. It’s just the latest tech for giving instructions to machines.