Dear Colleagues,

Just published this post on the new EU AI Act that will soon be coming into force and specifically for us, how it will impact AI-generated stock content.

https://brutallyhonestmicrostock.com/2024/01/26/new-eu-ai-act-and-its-impact-on-ai-generated-stock-content/

Alex

Lets start, the agencies aren't in the EU. I'm not either. But lets say, OK if they want to sell there, they will need to abide with those regulations.

Problems with the disclosure Publishing summaries of copyrighted data used for training.

1) Yes, we looked your image. So what?
2) We downloaded your image? But it wasn't used in the training
3) We used your image to train the machines.
4) There's no way to prove a one to one direct use of your image, in any generated results.
5) Images are not directly used to create new images, the machines learn and create new images from their training.

But aside from that nest of snags and questions.

Fair Use What is fair use. The courts need to decide that.

I look up a word in the dictionary and find the spelling and meaning and I use that knowledge to write a sentence. I publish the sentence. Did I steal the meaning or the word? I look at a photo and find the meaning and learn about how it's composed and the subject. I create my own image, using that knowledge, I didn't copy the image, I just learned from it.

AI does not copy images or re-use bits of them in the new images. It learns from what's in the images, the content and shapes, and creates an entirely new image.

Fair Use?

AI training is not stealing images. As much as I don't like AI and I'm not impressed by people with seven fingers and three arms or hands growing out of their ears and face, it's not directly using my images in the output product. AI / machine learning, is looking and learning.

The courts will decide that, not us or some political agreement.

Hi Pete,

(a) Actually, "ai training" is actually stealing images. One of the big "problems" they've had is figuring out how to remove "watermarks". Sure - they create "models" - but it is based off of theft.
(b) The algorithms (instructions) are not actually "learning". Just because the computer textbooks say it is "machine learning", doesn't actually mean the machine is "learning". It is not. For clarification - it is code a person wrote to look for "patterns" in order to create a "model representation" (i.e., compressed data structure), so instead of 100 gb of say pictures of "oranges", it could be say "represented" in only 0.5 gb, and then taking varying inputs (i.e., "red" orange, "blue" orange, etc) - to essentially 'blend' models to 'make' new stuff - based off of theft.
(c) The "courts" (i.e., like a 'tennis court') - simply (sadly right now) - basically do whatever someone with the most $$$ tells them to do. (Look @ the last 4 years - just follow the money to figure out what really went on).

The "ai" actually is directly using your images, as well as the images of others, and essentially creating composite images. It does not create "composite images" on the fly (it is not simple alpha blending) - but basically a more sophisticated form of theft - basically "pre-blending" the images in a "model" - and then using that "model" to generate new images. Theft is still theft. One does not need a "court" to decide if it is theft - when it is in fact theft.

Lets start, the agencies aren't in the EU. I'm not either. But lets say, OK if they want to sell there, they will need to abide with those regulations.

Problems with the disclosure Publishing summaries of copyrighted data used for training.

1) Yes, we looked your image. So what?
2) We downloaded your image? But it wasn't used in the training
3) We used your image to train the machines.
4) There's no way to prove a one to one direct use of your image, in any generated results.
5) Images are not directly used to create new images, the machines learn and create new images from their training.

But aside from that nest of snags and questions.

Fair Use What is fair use. The courts need to decide that.

I look up a word in the dictionary and find the spelling and meaning and I use that knowledge to write a sentence. I publish the sentence. Did I steal the meaning or the word? I look at a photo and find the meaning and learn about how it's composed and the subject. I create my own image, using that knowledge, I didn't copy the image, I just learned from it.

AI does not copy images or re-use bits of them in the new images. It learns from what's in the images, the content and shapes, and creates an entirely new image.

Fair Use?

AI training is not stealing images. As much as I don't like AI and I'm not impressed by people with seven fingers and three arms or hands growing out of their ears and face, it's not directly using my images in the output product. AI / machine learning, is looking and learning.

The courts will decide that, not us or some political agreement.

Hi Pete,

(a) Actually, "ai training" is actually stealing images.

In your opinion. I tried the detailed explanation, now the shortest version possible. Images are not directly used to create new images.

Yes, the courts will decide this and if it is Fair Use or not.

Your reasoning for why hands are wrong is one of the most laughable things I've read today. You clearly are in denial or don't understand how AI/Machine Learning works. "not enough examples" 

LAION, the nonprofit that put the database together, used web crawling tools to create datasets with more than 5 billion links to online images, which companies can then pull from to use as training data for their own AI models.

AI does not understand function or purpose, it only understands training. AI doesn't know how many fingers or how many legs on a chair, or how many legs on a horse. Not because it hasn't seen enough in 5 billion images, but because AI lacks reasoning. AI doesn't understand anatomy, only what it has seen. And since AI creates new images from training, it might have more or less fingers, based on the random samples it uses to create the new hand.

Hi Pete,

a) I do agree with you when you say that images are not "directly" used to create new images. That is correct. You may have misunderstood or misread what I wrote. But the images are stolen to create the initial model - in order to create a "model" - from which new images are then created.
b) The courts may "decide". If they decide to say that theft is not theft, it does not mean its not theft. One does not need a 3rd party "authority" to decide if something is wrong, when they know it is in fact wrong.
c) I actually do understand how "machine learning" works, in fact I even have a background in it (among other things). "AI" (at least the "ai" that is used now) - can't "reason", doesn't re-write its "code", etc. It is a set of algorithms that computer programmers have put together, to create a "model" (i.e., mathematical representation) of what say an "apple" is. The reason the "ai" machines need keywords (and shutterstock,etc just "gave" them, then gave most contributors a paltry some for their data) - was so it knew what words/phrases to associate scraped (stolen) images with. But it is - in fact - stealing.

Basically - the companies have no idea what a "cat" is. So they scrape (stealing) a bunch of random images. But they need to associate keywords with the images they scrape. The algorithm then essentially determines a pattern (i.e., say a "cat" has "lines" representing "whiskers", has a "round head", has two "triangle ears"). It then "compresses" this data (so instead of storing 100,000's of images, it simply says "a cat is an object which has whiskers, round head, and triangle ears") - which it deterimined soley by keywords/etc associated with it, and the scraped stolen data.

It's kind of like using winzip to zip a text file. You could have a 100MB "text" file, and then a compressed 1mb "zip" file (10000 times smaller). But the original "data" is still there, encoded. The latent diffusion model essentially just "blends" multiple text files together and then when you 'unzip' a file its decoding the file.

What is going on in the "AI" space is kind of like...

Someone "knows" someone "else" stealing a bag of apples from a street vendor, and drops several apples on the street. The person (group/organization) "stealing" the bag of apples justifies their actions saying it was "research", and they are just "studying pyschological effects" blah blah blah, and have "funding" from various sources... the person stealing the apple from the street is the "ai" company knowing that they were stolen apples, but pretends to play innocent saying "oh gosh darn golly gee, I just saw the apple lying there"...

If the person makes it into apple jelly, or apple vinegar, or whatever - they still knowingly took it knowing it was theft...

And what makes things even more interesting, is if you find out the person that "found" the apple - was the one that "funded" the person doing the "research"...

What "they" (most of the "ai" companies) are doing is theft, plain & simple. When your major "problem" is "watermarks" - and how to get rid of those "pesky things" - where do you think "watermarks" come from? Certainly not properly licensed content where the creator of those works has given explicit consent...

Quote from: SuperPhoto on January 27, 2024, 08:01

Hi Pete,

(a) Actually, "ai training" is actually stealing images.

In your opinion. I tried the detailed explanation, now the shortest version possible. Images are not directly used to create new images.

Yes, the courts will decide this and if it is Fair Use or not.

Your reasoning for why hands are wrong is one of the most laughable things I've read today. You clearly are in denial or don't understand how AI/Machine Learning works. "not enough examples" 

LAION, the nonprofit that put the database together, used web crawling tools to create datasets with more than 5 billion links to online images, which companies can then pull from to use as training data for their own AI models.

AI does not understand function or purpose, it only understands training. AI doesn't know how many fingers or how many legs on a chair, or how many legs on a horse. Not because it hasn't seen enough in 5 billion images, but because AI lacks reasoning. AI doesn't understand anatomy, only what it has seen. And since AI creates new images from training, it might have more or less fingers, based on the random samples it uses to create the new hand.

While I don't agree with SuperPhoto on everything (some of it sounds like a conspiracy to me), I have to agree with him that the AI image generators actually use real images to create new ones. This is done by morphing from many learned images.

In 99% of cases you don't notice it, especially with very generic subjects, of which an extremely large number have been used. Here, the morphing makes it unrecognizable and makes you believe that the AI is creating completely new images on its own.

But if you pick out special subjects, e.g. photorealistic wildlife shots of animals, of which there were obviously not that many used for the training, it sometimes happens (at least with Stable Diffusion) that you discover blurred copyright labels in the corner of the image.
So if the AI would generate complete new images than copyright symbols should not appear.
The AI image generators currently have no ability to abstract like a human and create new things by itself. It's only morphing of pictures.

@SuperPhoto
I would argue that a conspiracy on a political level happens not in every corner of the world. I don't know what it's like in the USA. I live in Germany and can't see any political leadership, strategy or conspiracy at all. The country is only living off the substance of the last few decades and is slowly going down the drain.

Quote from: Andrej.S. on January 28, 2024, 07:44

...
But if you pick out special subjects, e.g. photorealistic wildlife shots of animals, of which there were obviously not that many used for the training, it sometimes happens (at least with Stable Diffusion) that you discover blurred copyright labels in the corner of the image.
So if the AI would generate complete new images than copyright symbols should not appear.
The AI image generators currently have no ability to abstract like a human and create new things by itself. It's only morphing of pictures. ...

one source of the copyright  is not taken from 1 image but from the multiple copies of that image on free sites.

once again, AI doesn't 'morph' pictures - the billions of images are each broken into many small matrices and transformed before being saved. then images are created de novo, starting with a completely randomized 'image' & making many thousands of passes as that new image slowly emerges. you can get an idea of how this proceeds by watching the midjourney cevelopment

Well, it's not that easy as it looks at the first glance.
I took sometime ago a rough look at the papers of the models.

Where we can agree is that not the original images are "morphed", but the image informations as a stochastic distribution of pixels and patterns / noise.

This image information is linked to text information in vector spaces so that text inputs (prompts) can recursively reconstruct images using learned image informations (e.g. difussion models).

Whether the original images, which were used to train the AI, can be reproduced completely or at least very closely depends heavily on the calibration quality of the model.
If there is an overfitting, the original images can be reproduced very closely.

The latest version of Midjourney, for example, has this problem.
There was once a news tech article about that movie scenes from Marvel movies were reproduced almost completely by referencing the movies in prompts.

copyright watermarks are often added as a design element by the ai. same with signatures.

they are trained on watermarked images and images with signatures, so it thinks this is a normal part of pictures.

which again shows the ai has no intelligence and is just pixel mixer machine.

Exactly this. The AI model learned that the copyright signature is just a part of the image (some pixel patterns of the image), which is labeled with keywords in vector spaces. When you use these keywords in the prompt it can happen that the AI generator will reconstruct a part of the trained image with it's copyright signature.
It is therefore obvious that unlicensed material was used for the training.

I even had once a half watermark from a photo agency with the normal SDXL model from Stable Difussion.
I had tried to generate a woman practicing yoga in her living room.

I generally see a high risk with open source AI generators, where every private user can train their own models with their own content (see civitai community).
There is no guarantee that the user has actually used licensed material.

In the end, it's all about the extent to which works protected by copyright can be reconstructed.
That alone is the most important point.

If one wanted, one could optimize the calibration quality of every model so that no overfitting occurs. That would be one of the possible legal restrictions.
But of course, this would have a negative impact on quality.

I therefore see only small chances of a successful lawsuit against Midjourney or Stable Diffusion.
It will simply be passed on to the users that they themselves are responsible for the copyright infringements.

Except the machine learning vendors broke into the bookshop at night, copied every single word and it�s relationship to every other as positioned by artists then spewed them back out again in a new alternative for money.

The problem with hands is, not all hands are flat out, four fingers spread, and a thumb. Some are holding things, some are pointing, some are shaking hands or holding hands. At that point and add variations of how many angles or fingers showing, the AI doesn't understand what a hand consists of, since there's no way to know the exact situation, and how many fingers are supposed to be there? We get the monsters.

Hands are an extremely complex human organ. Extremely flexible, linked to non-verbal language that goes "hand in hand" with facial expressions and body gestures.

AI would therefore have to be taught anatomy, gestures combined with language and emotions.
This is roughly what a human learns over the course of decades through day-to-day communication and social context.

But for simple static AI image generation, you probably only need gigabytes of hand images. But for AI videos ... this might be a very long way.