https://openjur.de/u/2495651.html

The inability to place AI-generated images under copyright is often seen as one of its biggest drawbacks, if you want to call it that. This is because AI-generated works lack the necessary depth of creativity. Just humans are allowed to claim copyright for its work. However, for me, this is not an issue, as many other areas of art operate under similar conditions. For example, textures for games. And also jourmalists can sing a song about articles that cannot be copyrighted.

One important consideration is that some AI companies require licenses for using their software. The new video model, Hunyuan, for instance, does not have a valid license for the European Union. This means you cannot legally use the software in those countries. And, as a result, you could be forced to take down work created with it.

This brings me to the repeatedly false claims about AI, starting with the assertion that AI “steals.”

From a technical perspective, it is not possible for an image AI to steal images. There is not a single pixel stored in the AI weights. AI “invents” its images from a noise pattern generated randomly with each seed. There is no way to store 240 terabytes of data from the training data set LAION for example in the 8, 12, 16, or, at best, 24 GB of VRAM on a graphics card. And this is just one of the training data sets. Achieving such a feat would win you a Nobel Prize. It is also impossible for the system to download specific images at runtime. AI models like Stable diffusion function offline too, so there is no source from which they could “steal.”

Another proof lies in the results themselves. Even when you explicitly prompt for the Mona Lisa, the generated output will always differ from the original. AI cannot replicate anything pixel-perfect because it does not start with an image. It begins with a grayscale or diffuse image, a noise map. This is why it’s called a diffusion model. While learning it goes the other direction by the way. It turns the source images into noise. And learns with every step what makes the image. But that just as a sidenote.

What AI retains in its weights is the concept and idea, not the exact image. Of course, copyright can still be infringed if you use AI to create content that mimics something already copyrighted, like a Super Mario character. However, this is not the fault of the AI. You could achieve the same result by drawing Super Mario in Paint, creating it in Photoshop, or modeling it in 3D.

Another claim is that AI infringes copyright during training, as this is when it interacts with the image material from the training data. This argument assumes that the data is downloaded and used, but it isn’t. It is streamed. It gets cached, analyzed, and then immediately deleted. For copyright infringement to occur, parts or the entirety of the image would need to be used. This does not happen. AI simply analyzes the image, much like an artist walking through a museum and studying the old masters. This kind of learning is not prohibited.

Finally, there is the claim that AI uses training material that explicitly forbids such usage. Even in Europe, there is a fair use clause that allows data scraping of publicly accessible content. The assertion that AI companies engage in illegal scraping is simply incorrect. If this were true, we would see a flood of lawsuits. But we don’t. The few cases that have gone to court have failed. Moreover, modern AI companies typically respect requests to exclude certain data from training.

Summary, there is simply no legal aspect that forbids the usage of AI. Be it image, music creation or text models.