(C) Daily Kos
This story was originally published by Daily Kos and is unaltered.
. . . . . . . . . .



Stable Diffusion AI is a Thieving Thief Who Thieves [1]

['This Content Is Not Subject To Review Daily Kos Staff Prior To Publication.', 'Backgroundurl Avatar_Large', 'Nickname', 'Joined', 'Created_At', 'Story Count', 'N_Stories', 'Comment Count', 'N_Comments', 'Popular Tags']

Date: 2023-03-08

A new paper shows that image generating AIs can steal wholesale from their training sets:

Armed with new and existing tools, we search for data replication behavior in a range of diffusion models with different dataset properties. We show that for small and medium dataset sizes, replication happens frequently, while for a model trained on the large and diverse ImageNet dataset, replication seems undetectable. This latter finding may lead one to believe that replication is not a problem for large-scale models. However, the even larger Stable Diffusion model exhibits clear replication in various forms (Fig 1). Furthermore, we believe that the rate of content replication we identify in Stable Diffusion likely underestimates the true rate because the model is trained on a large 2B image split of LAION, but we only search for matches in the much smaller 12M “Aesthetics v2 6+” subset 2212.03860.pdf (arxiv.org)

Stable Diffusion would do things like take whole elements of a picture and put them into a generated image or use a background from one image in another from the training set or mix and match different elements in a generate damage -- things that we would let no human artist claim as original work. In addition, it would produce images in the style of an artist by effectively just changing a few elements, sometimes as few as a few pixels. Copying, in other words.

What was interesting about this, however, was that such copying was not inevitable, at least not at the wholesale image copying level. The ImageNET LDM model did not have this copying issue. Now, it seems ot require a large training set (all the smaller models they tested had copying issues) which raises its won set of ethical issues around the consent of artists, but it does highlight that the way these models produce works is a choice. The specific ethical issue of copying could be overcome, then. But the makers of the Stable Diffusion model did not do so, in part because there is no incentive for them to do so.

Right now, we treat artificial intelligence as if it is some special case, some fragile piece of magic that we risk destroying if so much as look at it sideways. That is complete nonsense. It is just code, like any other code. There is no intelligence in it. It has no magic, just a set of instructions. But because we have allowed ourselves to be conned into believing it is special and thus deserving of specifical treatment, there is no incentive for companies to produce models that are ethical instead of unethical. We get thieves.

Culture is incredibly important to the health of any society. That is why we have rules to support and nurture it. AI might be a net contributor to our culture if we can set proper boundaries for its use. But where those boundaries lie are for the entirety of society to decided, not just for the businesspeople looking to make a quick buck of lines of code. Generative AI is like any other tool -- it requires proper regulation to ensure its externalities are not just foisted on the rest of us to deal with why the benefits accrue to only the ones with the capital to fund it.

[END]
---
[1] Url: https://www.dailykos.com/stories/2023/3/8/2156894/-Stable-Diffusion-AI-is-a-Thieving-Thief-Who-Thieves

Published and (C) by Daily Kos
Content appears here under this condition or license: Site content may be used for any purpose without permission unless otherwise specified.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/dailykos/