Ben Werdmuller: There are some jaw-dropping infringements here, including an image where DALL-E apparently copies the entire Pixar universe from the single two-word prompt, "animated toys".

There are some jaw-dropping infringements here, including an image where DALL-E apparently copies the entire Pixar universe from the single two-word prompt, "animated toys".

Things are about to get a lot worse for Generative AIgarymarcus.substack.com

it would be ironic/iconic if AI in the stupidity and greed of it's designers functionally tenders IP useless for them

Sexually Suggestive

Labeled by Bluesky Moderation Service

Voracious media conglomerates vs unaccountable tech magnates

Just so people can see the problem: you don't need to name a copyrighted character to get DALL-E to produce an image of one.

"Oh, hello there meat bags"

Hey, it's C-3PX

The one thing the article gets wrong is that this is somehow hard to fix. They could easily take all the infringing and unlicensed material out if their training dataset and train a new model.

They don’t have a dataset if they do that. The vast majority of it is not in the public domain It’s also huge, so taking everything infringing out is a monumental task

Maybe they should have thought of that before they used a dataset of which the vast majority was copyrighted

Exactly, how can you base your business on relying on work from others but not paying them properly?! There is this stupid thought that everything must be free on the internet...

A) then they should develop a more sophisticated training method that needs very much less data B) they are already paying for huge amounts of human post-processing work. They could redirect that effort to non-infringing training set creation

If they knew how to do A, they would, because this way is expensive AF. But they don’t, so they’re going to die on the hill of “massive copyright infringement is fine when it makes us lots of money”

"But if you are taking away all my slaves, who is working on my plantations and generating revenue for me then?" 🤷

the “training a new model” part is, I think, still technologically unsolved (not straightforward to redo months of computation every time someone asks their data to be removed), but I’d argue that that’s their problem to solve, rather than shrugging and plowing ahead anyway

There are methods for concept editing (e.g. unified.baulab.info ) in diffusion models, but they won't be equivalent to not having pictures in the training set

Unified Concept Editing in Diffusion Modelsunified.baulab.info Addressing bias, copyright, and offensive content in diffusion models by directly calculating parameter changes.

The can train a new model with a smaller/alternative dataset in the se way they trained the model they have right now. Training a new model on new training data is always possible.

Mosaic has the CommonCanvas project - arxiv.org/pdf/2310.168... - which attempts to do this by only building off CC-BY-licensed images, which does make it harder to recreate exact likenesses, but does not make it entirely immune either

arxiv.org

Thought 3PO was giving me the finger here

Well, it could be worse.

I really think AI generated stuff - of any sort - needs to cite or provide general information on how it generated the content.

I can understand how this happens with pictures. Creating a picture needs a much wider definition of the request and the model will have a very limited corpus of information to work from. Text reproduction is another matter and calls into question my assumptions re the training.

To put it another way, a picture is a coherent whole, with elements that can be reused but mostly rely on the context of the picture. A text, and the relationships between the words, is far more instructive and, in my view, highly unlikely to be reproduced by a prompt. Yet here we are.

Two thoughts. Yes, the limits of AI is that it is not capable of coherent original thought. I fully expect the AI industry to ultimately end up challenging copyright laws & licencing as they stand: AI Vs What is original thought anyway...

The appeal: isn't this all in the public domain, who are you the thought police. The appeal judge: yes, yes we are & moreover you have failed to show any evidence of actual intelligent thought...

The appeal, appeal case: piracy is a long held tradition that is endemic across the globe, it is unrealistic to expect AI not to follow the natural laws of society. The appeal judge: lol... That is not the case before me & the not my problem. Perhaps your AI could explain it to you. Dismissed...

Copyright laws will be used to do something morally good for once 🙏

This is good news. Time to throw out the offending algorithms and start all over. AI in its current iteration is appalling.

The thing is, I think this _could_ be fixed, in that metrics could be found for distance in parameter space from copyright materials could be identified and if too close, the output could be blocked.

But on the other hand that would render the model much less functional and maybe that should tell us something

Well, that seems potentially problematic.

Post