Thread by @wickedinsignia

Here's a megathread of responses to the many ill-conceived arguments regarding AI. I hope it serves as a resource to skeptics and artists:

"This is the democratization of art!"
It cannot be, since it occurred without existing artists' vote or input. The technology is built directly on the back of artists' work without permission or compensation, and its adverse effects on their livelihoods are left to run rampant.

"This is just like how humans learn!"
An image-based AI is not a person. It does not have human-level comprehension and cannot be inspired in the same way. This has been acknowledged by neuroscientists such as Henning Beck and by deep learning experts:

https://twitter.com/fchollet/status/1563153088470749196?s=20&t=OHaLwZKu7OBZeGt1_qROBQ

It's also worth point out that a human artist is not an instant online service developed, owned and maintained by a company. Money given to artists lands in the pockets of many thousands of independent laborers. Money given to AI lands in the hands of a few select companies.

"You're trying to stop progress."
Progress would be taking the livelihoods and rights of artists into account and developing technology that respects common ethics. Current image AI does non of that. Data and copyright exploitation is not progress.

"You're a Luddite."
The ethical concerns around AI extend far beyond the concept of a new working method. They raise important questions around abuse and exploitation of data.
You'll also want to become acquainted with what the Luddites actually stood for:

https://twitter.com/mollycrabapple/status/1602299750367232000?s=20&t=OHaLwZKu7OBZeGt1_qROBQ

"You're gatekeeping!"
Art is already easily accessible to all. Regulating a technology that's built directly on top of our work is not gatekeeping and exploitation of labor is not a right. Artists have as much of a right to protect their work as you have to make art.

"AI does not breach copyright."
Overfitting would strongly disagree with you. A diffusion model AI could spit out an image identical to one in its training data at any moment. All diffusion-based models are prone to this and StabilityAI admits to it.
wandb.ai/wandb_gen/audi…

"AI doesn't contain artists' images directly"
Image AI is trained on a dataset of images and it's true that the image data is not retained 1:1 in the final algorithm.
However, the quality of the algorithm's output is almost entirely dependent on the contents of that dataset.

If you took all copyright images out of available datasets, the quality of the resulting image AI would be exponentially reduced.
Their capabilities are defined by the presence of work the developers do not own. Artists' work is used directly in the developmental phase...

...without permission. Considering that many of these models are profitable services, that also means a commercial product is being created directly through the utilization of content with variable licensing, much of which disallows use in a commercial product.

As mentioned before, Overfitting is a thing. This means you have a commercial product capable of reproducing copyright works that were directly involved in its development. Not to mention that it would not be as capable and therefore profitable without the presence of said work.

"Stable Diffusion is open-source. What's the problem?"
Artists' work could be designated with any variety of licenses. Open-source is subject to a license, which in many cases will not match the license artists apply to their work. Developers have no right to apply...

...whatever license best benefits them to content they don't own.
An artist's work should be subject to the licenses and use they impose upon it, not freely re-allocated at the whim of a company.
Artists' work in the developmental phase of AI software raises many legal questions.

"Your art is on the internet, so it's ours to use."
That's a fundamental misunderstanding of law. DMCAs, takedowns, and copyright exist for a reason. Even if current law contains oversights that allow scraping, that is subject to adjustment based on a rapidly changing landscape.

"This is out now, you can't take it back."
These tools were made available to the public by companies which can and should be held accountable for their actions. Technology isn't an unstoppable force devoid of regulation, and actions from both users and devs can be mitigated.

Technology should not be treated as something that is constantly thrust upon us, leaving us powerless to stop or alter it. We have the power to provide input and adjustment to emerging tech and are now exercising that at a most crucial time.

"So you're just anti-AI?"
No. Skeptics are fighting for an environment where AI development cannot exploit work/data we believe it never had any right to use and where the adverse effects on many people's livelihoods are mitigated. If AI was built and introduced...

...responsibly, there would not be an issue. All technology requires friction and adjustment to benefit the lives of everyone that it affects, and all technology is subject to adverse use by those with ill-intent or by belligerent self-interested idealists.

"It's just like when the camera was invented!"
Image AI is not a camera, nor akin to any other invention. The practical similarities are extremely meager in any case. AI brings with it a slew of unique social and ethical concerns that the camera and prior inventions did not.

This type of argument is simply reductive. There may be a loose similarity to previous events (the camera, the printing press etc.) but the implications of the technology and the way it's built are entirely different. It does not invalidate the criticisms being expressed now.

"Anyone can save images to disk. What's wrong with data scraping?"
Technically, nothing! Scraping data is not a problematic action on its own. Sometimes it may go against platform ToS (in Pinterest's case) but it's not problematic to save a ton of publicly available data.

It's how the data is then used that's problematic. Utilizing a huge dataset of copyright works (such as Laion5B) in the direct development of software raises many unanswered questions re: legality. Many artists believe this use is unethical and that the laws should reflect this.

“You literally draw fanart.”
Yes, fanart is technically an act of copyright abuse. Companies treat it as an exemption because it does more to advertise for their brand than profit unfairly from it. Also, no fan artist is claiming they came up with Sonic the Hedgehog or Mario.

If someone attempted to mass-produce their fanart (such as starting a merchandising business at a large scale) they would find themselves quickly slapped with a C&D.
AI imagery that infringes copyright usually takes credit for work by small creators/artists, not a renowned brand.