AI Data Laundering

https://waxy.org/2022/09/ai-data-laundering-how-academic-and-nonprofit-researchers-shield-tech-companies-from-accountability/

by marceloabsousa • 4 years ago

304 113 4 years ago

waxy.org

dkural• 4 years ago

This reminds me of the Jedi Mind trick of Uber of waving a smartphone to argue that labor & other laws all of a sudden don't apply to them, to the detriment of the public that'll now shoulder the costs.

moyix• 4 years ago

The Authors Guild v Google decision about Google Books seems relevant:

In late 2013, after the class action status was challenged, the District Court granted summary judgement in favor of Google, dismissing the lawsuit and affirming the Google Books project met all legal requirements for fair use. The Second Circuit Court of Appeal upheld the District Court's summary judgement in October 2015, ruling Google's "project provides a public service without violating intellectual property law." The U.S. Supreme Court subsequently denied a petition to hear the case.

[...]

The court's summary of its opinion is:

[...]

Google’s unauthorized digitizing of copyright-protected works, creation of a search functionality, and display of snippets from those works are non-infringing fair uses. The purpose of the copying is highly transformative, the public display of text is limited, and the revelations do not provide a significant market substitute for the protected aspects of the originals. Google’s commercial nature and profit motivation do not justify denial of fair use.

https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....

This doesn't touch on the ethics of course – at minimum I think allowing people to exclude themselves or their work from a dataset is necessary.

authpor4• 4 years ago

I think allowing people to exclude themselves or their work from a dataset is necessary.

or they could open it all up for everybody and stop protecting the rights of death people (authors dead less then 70 years ago)

then again, that will make the publishers starve... but why pretend publishing corporations need food?

ad404b8a372f2b94• 4 years ago

This is larger than publishers, this is every artist, film-maker, photographer, every writer, every engineer, anybody who has ever written or created something and shared it publicly is liable to have their work assimilated and an infinite amount of derivatives produced with no control over how they're used and by whom.

Comment generated with gpt-neox prompt: Comment about AI and data collection and generation and its pitfalls, expressing concern, emphasis on professions, emphasis on automation, written by Stephen King, creative writing, award winning, trending on reddit, trending on hacker news, written by Greg Rutkowski, written by Zola, written by Voltaire, written by authpor, written by moyix.

(Just kidding, it wasn't AI generated but you see my point.)

TigeriusKirk• 4 years ago

My personal ideal outcome is that there's no opting out of having your intellectual output included in the training, but the resulting model is as a result available freely to the public.

In my utopia, the end results are models containing the sum total of human output, available to everyone.

What I think is unconscionable is training the models on public works and then retaining them exclusively for private use.

visarga• 4 years ago

For the first time there is a chance for Mickey Mouse to be free, I mean "In-the-style-of-Mickey-Mouse", his new name. When did we ever get such a chance for information freedom?

noduerme3• 4 years ago

Why pretend that other corporations that vacuum up content and repackage it have rights to resell art that you want to strip from the original publishers? At least the publishers actually made a contract with the artists.

echelon3• 4 years ago

Do we allow artists to withhold their works from the minds of eager, learning children? [1]

Tell me how ML is different than the mind of a toddler ravenous for new information.

For every billion dollar start-up using data at scale, there are tens of thousands more researchers and hobbyists doing the exact same, producing wonderful results and advances.

If we stop this growth dead in the tracks, other countries more willing to look past the IP laws will jump ahead. And if Stability locks away their secret sauce, some new party will come and give away the keys to the kingdom yet again.

You can't block the signal. Except, of course, by legislating against it in some Luddite hope we can prevent the future from happening.

Instead of worrying careers will end, we should look at this as being the end of specialization. No longer do we need to pay 20,000 hours to learn one thing to the exclusion of all others we would like to try. Now we'll be able to clearly articulate ourselves with art, music, poetry. We'll become powerful beings of thought and expression.

Humans aren't the end or the peak of evolution. We should be excited to watch this unfold.

[1] Maybe Disney would like you to pay more for a premium learning plan for your child, but thankfully that's not (yet) possible.

nightski2• 4 years ago

Well a toddler isn’t making money off the information they are absorbing for one. If these are open to the public models that is one thing. But no, these are proprietary models whose sole purpose is to make money for large corporations.

9wzYQbTYsAIc17• 4 years ago

Tell me how ML is different than the mind of a toddler ravenous for new information.

If a person published a work that clearly plagiarized or violated a patent, that person would be open to legal action.

I’m all for systemic change, but uses like this may end up having a chilling effect on human-created work.

greysphere2• 4 years ago

Most machine learning is assigning weights in a chain of matrix multiplications and normalization functions.

There is no known experimentally verifyiable model of toddlers' brains, let alone one based on matrix multiplication and normalization. Developing such a model would be a noteworthy achievement.

Therefore these are different.

numpad03• 4 years ago

The criticism is that AI works are not transformative, but are recognizable “regurgitation” of training set.

It’s not that AIs are too good. They look like crude knockoff products to trained eyes. And crude knockoffs are usually considered bad things.

_jal• 4 years ago

Tell me how ML is different than the mind of a toddler ravenous for new information.

Well, I can't keep a toddler in a data center, pumping out work on demand. Or copyright it and limit who it chooses to work for when it grows up.

For instance.

boppo14• 4 years ago

No longer do we need to pay 20,000 hours to learn one thing to the exclusion of all others we would like to try. Now we'll be able to clearly articulate ourselves with art, music, poetry. We'll become powerful beings of thought and expression.

I'm a 20000 hours person. Knowing what I know about what I do, it's real sad to see someone misunderstand what goes into creativity this egregiously. Prompt engineering is such an unbelievably watered down "version" of making a painting. It's like writing a page, or even a folder! of bullet points and handing it to a ghostwriter, then telling them "put the end result between Shakespeare and Poe".

That's not unleashing your creative voice. Unleashing your voice and acquiring technical skills in a chosen field are the same. If you endlessly mixed all the prior classical works, it doesn't matter how you weight them, it won't spit out Mozart. You're stuck in the gamut of the model, between the maxima points of each artist.

It's an incredible tool to generate stuff quickly, and to some extent it will help artists whose work depends on quantity over quality.

trention2• 4 years ago

Tell me how ML is different than the mind of a toddler ravenous for new information.

The toddler is human. AIs are not humans.

It's a human right to learn. Non-humans don't (and shouldn't) have human rights.

Humans aren't the end or the peak of evolution. We should be excited to watch this unfold.

Spoken like a true evolutionary loser.

9wzYQbTYsAIc17• 4 years ago

the revelations do not provide a significant market substitute for the protected aspects of the originals

It does seem like generative AI systems provide a significant market substitute, so this ruling probably wouldn’t apply, in court.

edit: see https://news.ycombinator.com/item?id=33194623 for some initial thoughts on how this problem (and others) could be rectified.

For example, with a database of protected works and self-censorship algorithms for generative AI systems, conscientiously objecting creatives could have a mechanism for excluding their works.

tpmoney2• 4 years ago

A substitute for what though? Copyright law is only concerned with substituting the work under copyright. That is to say, the consideration is whether the infringing aspects of the secondary work would alter the demand and market for the work being infringed.

In all the talk about AI data laundering there really hasn't been any indication that the AI generated item substitutes for the item it's alleged to infringe on. Substituting for a whole profession and its practitioners doesn't enter into the concerns of copyright law. There might be some argument that it should (to "promote the progress of science and useful arts" as it were), but copyright law to my knowledge hasn't been used to prevent new tech from putting professionals as a whole out of business.

VanTheBrand5• 4 years ago

I would argue (as the court did) that google's use is transformative because the end result "book search" is in a different marketplace from "books." The end result / output of these generative AI systems trained on stock media and art is..."stock media and art."

That's kind of what this whole article is about. Just training the systems in research is arguably fair use but creating the entire pipeline might not be and the "loophole" here is trying to claim no responsibility for the training at the center of it because that was technically done by a 3rd party (...funded by the final creator of the full entire pipeline.)

pavlov• 4 years ago

The court’s summary also mentions this aspect of differing marketplaces:

“… the revelations [i.e. the information served by Google Book Search] do not provide a significant market substitute for the protected aspects of the originals.”

This doesn’t apply to AI image generators which are clearly a “market substitute” for the protected originals used to train the system. For this reason I’d expect someone like Getty to want to revisit Authors Guild v Google sooner rather than later.

russellbeattie• 4 years ago

An important part of the opinion (on the wiki page you linked to) is completely missing in the case of AI datasets:

It generates new audiences and creates new sources of income for authors and publishers.

This is definitely not the case for artists and photographers, who don't benefit at all from the transformative nature of the AI output, and in fact are significantly harmed since it dilutes the uniqueness of their work by allowing anyone to imitate their style. Though to my knowledge "style" isn't protected by copyright - only trademark - I can't imagine there won't be lawsuits about this in the future.

That one artist who complained that people can't find his original work online now because of so many imitated pics is definitely exhibit A in terms of direct harm.

dangerface• 4 years ago

Google’s unauthorized digitizing of copyright-protected works, creation of a search functionality, and display of snippets from those works are non-infringing fair uses. The purpose of the copying is highly transformative, the public display of text is limited, and the revelations do not provide a significant market substitute for the protected aspects of the originals.

So is digitizing a copyright vhs and hosting it via torrents also fair use? Its transformative, the public display of the video is limited, there is no market for vhs.

I don't get it whats the difference other than Google having deeper pockets than me?

killjoywashere• 4 years ago

It’s currently unclear if training deep learning models on copyrighted material is a form of infringement

What? It's clearly a derived work.

alar44• 4 years ago

Alright, then every piece of music you've ever heard is also a derived work. Unless the composer grew up in a void.

kevingadd• 4 years ago

I think you're trying to be ridiculous, but lawsuits of that sort do come up periodically - accusations of composition theft that in practice are probably not theft and just the fact that there aren't THAT many unique note progressions you can put into a song, and it's not that odd for a composer to accidentally imitate something they heard a long time ago when picking one.

wodenokoto• 4 years ago

I'm pretty sure I can count the number of words in Harry Potter without breaking copyright law.

It is absolutely not clear when statistical models stops counting ngrams and starts making a derived works.

tiborsaas2• 4 years ago

You can also read the HP series and write summaries and reviews about each book as wodenokoto. You can probably create HP looking artwork and write stories that could fit into the HP universe. You can't call any of your work HP. This sounds obvious, but if it's done by a machine then some people think it's a different question.

I can write code to get a list of characters in the book, get their page numbers analysed and draw graphs to help me create my own version. Am I breaking copyright laws? Most likely not.

It's a truly grey area which lawmakers never saw coming.

I believe if events unfold well we'll see and treat AI tools to be like sharp knives eventually. It will be up to the user what they do with it.

RosanaAnaDana• 4 years ago

The horse seems well out of the gate.

VanTheBrand5• 4 years ago

the horse is out of the gate on photocopiers and before them printing presses but that doesn't make using them without the rights to what you are copying legal

pixl97• 4 years ago

I would go as far to say a vast amount of our creative economic output would cease tomorrow if we had strict 'right to read' copyright enforcement. You can also thank Disney for extending copyright beyond all reason.

learndeeply• 4 years ago

But then Meta is using those academic non-commercial datasets to train a model, presumably for future commercial use in their products. Weird, right?

This is a very strong and likely inaccurate presumption.

brrrrrm• 4 years ago

yep. This class of fallacy has its own wiki article: https://en.wikipedia.org/wiki/Appeal_to_probability

nerdponx• 4 years ago

Is it? Maybe they have their own internal version they are using, but who's to say they aren't fine tuning the model and applying it somewhere?

Havoc• 4 years ago

The whole thing is a mess but frankly i doubt this genie can be put back in the bottle

9wzYQbTYsAIc17• 4 years ago

I think it is an a priori fact that the cat is out of the bag.

The existing publicly available datasettes, algorithms, and weighted models certainly should be expected to be permanently in the hands of some non-law-abiding parties, at this point.

I think that it will be important to ensure that we have symmetric information, going forward, otherwise trying to put the genie back in the bottle may just end up further disadvantaging those that try to follow the rules.

ROTMetro3• 4 years ago

...said the music industry about samplers in the 1990s.

gfd• 4 years ago

Was this term coined on HN? I remember first seeing it (used in an AI context) from this 2019 comment under "Cool stuff that's still completely unregulated": https://news.ycombinator.com/item?id=21167689

Most of the predictions in that first comment came true.

lachlan_gray• 4 years ago

William Gibson mentions data laundering as an illicit activity in the Neuromancer books! It’s plausible that the phrase itself was coined there

bo1024• 4 years ago

The Flickr example is wild. How was nobody sued for that!?

noduerme3• 4 years ago

I've got a couple examples of Stable Diffusion replicating watermarks along with similar swatches of imagery into scenes from the same prompt [1]. A single case of this should be enough to file a massive lawsuit if the art were recognizable to the creator.

[1] https://news.ycombinator.com/item?id=33061707

danielbln2• 4 years ago

The model learns all attributes of the images it's trained on, including that some have a watermark. The fact that it generates a watermark in some images doesn't mean that that is a 1:1 image from the training set, it just means to the model some images seem to have a watermark, so it will add it sometimes. Often you can just add "no watermark" (or add it as a negative prompt with some weights) and re-use the same seed to get the same image without the watermark.

ROTMetro3• 4 years ago

No, it means that it is reproducing the original work and is not producing a new original work. It is basically a really fancy Instagram lense, but it is still 100% derived from the underlying works and therefor derivative instead of a newly created non-derivative work.

noduerme3• 4 years ago

It may or may not be a 1:1 image, but I think it's significant that in both cases, with different seeds, what is directly behind / right of the watermark is a pretty similar building with different distortions applied to it. I'm not sure what the difference is between "learning" from a particular image and encoding that image with a lot of compression, when in either case the usage more or less reliably reconstructs the image algorithmically.

If I have a photographic memory and I memorize the Coca Cola logo and then draw it into a commercial work by decoding the firing of my neurons into muscle movements, the storage and retrieval method I used has no bearing on whether I infringed on their copyright.

nojvek• 4 years ago

Big Tech has really big datasets esp Google. With YouTube, Photos, Music, Gmail, Docs, Maps, Books, Waymo, Search … they have giant multimodal datasets that capture essence of all human knowledge. They have 10+ products with more a billion users creating data for them.

If Google Brain/DeepMind were to crack AGI, it would make Google/Alphabet crazy rich at the detriment of millions of YouTubers, Book authors, musicians, drivers.

AI will concentrate power and wealth to fewer individuals.

mirker• 4 years ago

Ads companies getting rich off of AGI seems a bit sensational when they’re already getting rich off of the boring type of AI. They’ve already gotten rich indexing the web and all the data we have years ago.

patcon• 4 years ago

Not sure laundering it the right term.

Laundering private things through the commons feels not as shady as laundering in private networks. The commons benefits too.

It's more like open source that money laundering

krab• 4 years ago

Are we heading towards voiding most of current copyrights or is there a way out of this mess with another patch to the laws?

theGnuMe• 4 years ago

It’s definitely fair use. One question I have though is Mickey Mouse protected by copyright or trademark or ? I assume someone other than Disney can’t sell mickeys likeness or is that wrong in art? And if the AI makes a movie?

AI Data Laundering

Related Stories