Show transcript
Screenshot of a tumblr post by hbmmaster:
the framing of generative ai as ātheftā in popular discourse has really set us back so far like not only should we not consider copyright infringement theft we shouldnāt even consider generative ai copyright infringement
who do you think benefits from redefining ātheftā to include āmaking something indirectly derivative of something created by someone elseā? because I can assure you itās not artists
okay Iām going to mute this post, Iāll just say,
if your gut reaction to this is that you think this is a pro-ai post, that you think ānot theftā means ānot badā, I want you to think very carefully about what exactly ātheftā is to you and what it is about ai that you consider āstealingā.
do you also consider other derivative works to be āstealingā? (fanfiction, youtube poops, gifsets) if not, why not? whatās the difference? because if the difference is actually just āwell itās fine when a person does itā then you really should try to find a better way to articulate the problems you have with ai than just saying itās āstealing from artistsā.
I dislike ai too, Iām probably on your side. I just want people to stop shooting themselves in the foot by making anti-ai arguments that have broader anti-art implications. I believe in you. you can come up with a better argument than just calling it ātheftā.
This is interesting. I agree that stealing isnāt the right category. Copyright infringement may be, but there needs to be a more specific question we are exploring.
Is it acceptable to make programmatic transformations of copyrighted source material without the copyright holderās permission for your own work?
Is it acceptable to build a product which contains the copyrighted works of others without their permission? Is it different if the works contained in the product are programmatically transformed prior to distribution?
Should the copyright holders be compensated for this? Is their permission necessary?
The same questions apply to the use of someoneās voice or likeness in products or works.
Somebody correct me if Iām wrong, but my understanding of how image generation models and training them works is that the end product, in fact, does not contain any copyrighted material or any transformation of that copyrighted material. The training process refines a set of numbers in the model, But those numbers canāt really be considered a transformation of the input.
To preface what Iām about to say, LLMs and image models are absolutely not intelligent, and itās fucking stupid that theyāre called AI at all. However, if you look at somebodyās art and learn from it, you donāt contain a copyrighted piece of their work in your head or a transformation of that copyrighted work. Youāve just refined your internal computers knowledge and understanding of the work, I believe the way image models are trained could be compared to that.
the generated product absolutely contains elements of the things it copied from. imagine the difference between someone making a piece of art that is heavily inspired by someone elseās work VS directly tracing the original and passing it off as entirely yours
I understand thatās how you think of it, but Iām talking about the technology itself. There is absolutely no copy of the original work, in the sense of ones and zeros.
The image generation model itself does not contain any data at all that is any of the work it was trained on, so the output of the model canāt be considered copyrighted work.
Yes, you can train models to copy artistsā styles or work, but itās not like tracing the image at all. Your comparison is completely wrong. It is a completely unique image that is generated off of the model itself, because the model itself does not contain any of the original work.
This is generally correct, though diffusion models and GPTs work in totally different ways. Assuming an entity had lawful access to the image in the first place, nothing that persists in a trained diffusion model can be realistically considered to be a copy of any particular training image by anyone who knows wtf theyāre talking about.
The magic word here is transformative. If your use of source material is minimal and distinct, thatās fair use.
If a 4 GB model contains the billion works it was trained on - it contains four bytes of each.
What the model does can be wildly different from any particular input.
Using peoples work, and math to make predictions is not transformative. Human creations are transformative.
Any transformation is transformative.