When training a model you are deriving a function that takes some input and produces an output. The issue with copyright and licensing here is that a copy is made and reproduced numerous times when training.
The model is not walking around a museum where it is an authorized viewing. It is not a being learning a skill. It is a function.
The further issue is that it may output material that competes with the original. So you may have copyright violation in distribution of the dataset or a model's output.
> a copy is made and reproduced numerous times when training.
Casually browsing the web creates millions of copies of what are likely the same images and text that models are trained on. Computers cannot move information, they can only copy it and delete the original. Splitting hairs over the semantics of what it means to "copy" isn't a strong argument.
> where it is an authorized viewing
What exactly is an unauthorized viewing of a publicly accessible piece of content online that has been hyperlinked to? If we assume things like robots.txt are respected, what makes the access of that data improper?
> it may output material that competes with the original
An art student could create a forgery. I could craft for myself a replica of a luxury bag. But that's not a crime unless it's done with the intention of deceiving someone or profiting from the work. Intent, after all, is nine tenths of the law.
It's an important right that you should be able to do and create things, even if the sale or distribution of the outputs of those things are prohibited. The ability for a model to produce content which couldn't be distributed shouldn't preempt its existence.
> So you may have copyright violation in distribution of the dataset or a model's output
And neither of those things are the act of training or distributing the model itself!
What makes the use improper? Licenses. Terms of service. Mostly licenses though. For example, all the images on Flickr that were uploaded under Creative Commons licenses (e.g. non-commercial) have now been used in a commercial capacity by a company to create and sell a product.
Similarly, code is on Github with specific licenses with specific terms. Copilot is a derivative work of that code, the license terms of that code (e.g. GPL, non-commercial) should extend to the new function that was derived from it.
The reason I mention competition with the original is the fair use test (USA). When courts decide whether something is fair use they consider a few aspects. Two important ones are whether it is commercial, and whether it is a substitute for the original. When art models output something in the style of a living artist, it is essentially a direct substitute for that person.
Sure, I can make a shirt with Spider Man on it and give it to my brother, but if a company were to use what I made or I tried to sell it, I would expect a cease and desist from Disney.
Training the model may very well be a copyright issue. The images have been copied, they are being used. Whether that falls under fair use will likely be determined on a case by case basis in court. I do not believe closed commercial models like Copilot or Dall-e will pass a fair use test.
There is a lot of money involved here though, so we will need to wait for years before we have answers.
1. https://www.theguardian.com/technology/2012/sep/11/minnesota...
This is not model training.
> Copilot is a derivative work of that code, the license terms of that code (e.g. GPL, non-commercial) should extend to the new function that was derived from it.
But the very act of training copilot is not problematic. And in fact, if GitHub never did anything with Copilot, the physical act of training the model is not problematic at all. And that's what at issue here. How Copilot is used is orthogonal to the article.
> Sure, I can make a shirt with Spider Man on it and give it to my brother, but if a company were to use what I made or I tried to sell it, I would expect a cease and desist from Disney.
Yes. And training the model isn't the part where you sell it. It's the part where you make it.
> Training the model may very well be a copyright issue. The images have been copied, they are being used.
What do you think "being used" means here? If I work for a company and download a bunch of text and save it to a flash drive, have I violated copyright? Of course not. If I put that data in a spreadsheet, is it copyright infringement? Of course not. If I use Excel formulas on that text is it infringement? Still no.
And so how can you claim in any way that the creation of a model is anything more than aggregating freely available information?
I don't disagree with you about the use of a model. But training the model is just taking some information and running code against it. That's what's important here.
How's that any different from what happens inside a human's brain when learning?
> The model is not walking around a museum where it is an authorized viewing.
The training data could well be from an online museum. And the idea that viewing something public has to be "authorized" is very insidious.
> The further issue is that it may output material that competes with the original.
So might a human student.
I have made no mention of things being authorized in public. In the US you are allowed to take a photo of anything you want in public. These models are not being trained on datasets collected wholly in public though, it is very insidious to suggest that they are.
The internet is not "the public". It is a series of digital properties that define terms for interacting with them. Now, a lot of material is publicly accessible online, but that does not mean that it is not still governed by copyright. For example, my code on Github is publicly accessible, but that doesn't mean you can disregard the license.
If you use this copyrighted material to produce a product for commercial gain you will likely face a fair use test in court. If you use it for a non-commercial cause with public benefit you could probably pass that fair use test. Open source will do very well because of this.
The model is not a human though, and very often these are not "public" works that it is trained on.
So is a human mind.
> In the US you are allowed to take a photo of anything you want in public. These models are not being trained on datasets collected wholly in public though, it is very insidious to suggest that they are.
How so? What non-public training data are they using, and why does it matter?
> The internet is not "the public". It is a series of digital properties that define terms for interacting with them. Now, a lot of material is publicly accessible online, but that does not mean that it is not still governed by copyright. For example, my code on Github is publicly accessible, but that doesn't mean you can disregard the license.
It does mean you can read the code and learn from it without concern for the license (morally, if not legally).
>How's that any different from what happens inside a human's brain when learning?
I don't know, nor does anyone else. So let me ask you - how is that the same as what happens inside a human's brain when learning?
We don't know the details. But it's pretty implausible that the process of learning wouldn't involve the brain having some representation of the thing it's learning, or wouldn't involve repeatedly "copying" that representation. Every way we know of processing data works like that. (OK, there are theoretical notions of reversible computation - but it's more complex and less effective than the regular kind, so it seems very unlikely the brain would operate that way)
And a human who has learned to perform a task has certainly "derived a function that takes some input and produces an output".
I think you can easily make a stronger statement:
We do know that art students spend many hours literally tracing other images in order to learn to draw. We do know that repetition is how the brain improves over time.
"Learn to draw better by copying." - https://www.adobe.com/creativecloud/illustration/discover/le...
Based on that, seems pretty clear to me that the other commenters here would agree (regardless what the brain does internally) that at a minimum, art students are violating copyright many, many, times in order to learn.
A few notable differences:
1. Scale: a single art student can't view millions of works in a week.
2. Duplication: a single art student's brain can't be cloned or downloaded into another art student's brain.
3. Speed: a single art student cannot draw or paint thousands of images in a week.
4. Ownership: the software is likely owned and controlled by a large corporation, while the art student (hopefully) isn't.
That was the point.
Whether or not there is copyright infringement going on (or whether copyright law is an appropriate regulatory framework for ML models), I frequently see claims like "ML training is no different from what humans do" repeated, in spite of its incorrectness.
There are many examples. One is face recognition - clearly it is acceptable for individual humans to do this at small scale, but systematic identification of everyone on a street or in a stadium has different implications for society.
(In this case it almost doesn't matter whether the surveillance is performed by humans or machines - it's the scale and systematic nature that changes the equation.)
The AI is not a human, but what you are doing is the same thing, if you claim the output as your work because you wrote the prompt.
Well, usually.
Sent from my iPhone.
Edit: and the browser have the same page to you and me. I bet you look at everything and think about how you can make lazy money with it.
To determine if an art student and DALL-E really are the same, despite their very obvious difference (one has arms and is part of a net of social relations while the other is intellectual property), will take some actual arguments which I presume you of course had planned to provide in a second comment from the start.
Just a rule of thumb.
If a human or AI reproduces a Van Gogh painting or derivative, it's not worth anything on the market.
Only original pieces, by a human artist, has value. A Van Gogh painting is worth millions of dollars only because it was created by Van Gogh.
That's the way I see AI-generated work going in the long run. The artists who have distinctive styles popular for image generation have seen a huge surge in attention, which I suspect will translate into making their original work more valuable.
A student learning from other artists is still limited in their output to human-scale - they must physically create the new thing. An AI model is not - the difference between a student learning from an artist and an AI model doing so is the AI model can flood the market with knockoffs at a magnitude the student cannot meet. Similarly, the AI model can simultaneously learn from and mimic the entirety of the art community, where the student has to focus and take time.
If this weren't capitalism - if artists weren't literally reliant on their art to eat, and if the market captured by the AI model didn't inevitably consolidate wealth - then we might be able to ignore that, but we do, and we can't ignore the economic effects when we consider scale like this.
I'm just sitting here wondering why it is even relevant whether the "AI" is "copying", "learning", "thinking", or whatever, why is any of that important? Does AI have human rights? Well, perhaps in a couple hundred years, if humanity manages not to self-extinguish by then.
It's not like you can sue AI if you think it plagiarized your work, no. Obviously not, so why the hell are we discussing that? "AI" is just a piece of software, a tool, it doesn't matter what it's doing, what matters is what the user is doing, the fact of the matter is that these multi-billionaire corporations are taking everyone's honest work, putting it into a computer, and selling the output. They didn't do any "learning", they just used your data and made money out of it, it isn't a stretch to say they simply sold your work.
EDIT: Perhaps one day the day AI will have human rights, make its own money, and pay bills. That will be the day any of this nonsensical discussion will be anything but useless.
And then there's the tens of thousands of people training models and making them freely available to everyone. What I fear most is that regulations introduced "to stop" the multi-billionaire corporations will in fact make sure they're the only ones with the resources to comply with the regulations.
In general I don't think this is the case, assuming you mean generations output from popular text-to-image models.
For DALL-E 2: I've never seen anyone able to provide a link of supposed copying. Even if you specifically ask it for some prominent work, you get a rendition not particularly closer than what a human artist could do: https://i.imgur.com/TEXXZ4a.png
For Stable Diffusion: it's true that Google did manage, by generating hundreds of millions of images using captions of the most-duped training images and attempting techniques like selecting by CLIP embeddings, to get 109 "near-copies of training examples". But I'd speculate, particularly if you're using the model normally and not peeking inside to intentionally try to get it to regurgitate, that this is still probably lower than the human baseline rate of intentional/accidental copying.
LLM’s recreating training material causes real issues such as Google’s dealing with PII leaks: https://ai.googleblog.com/2020/12/privacy-considerations-in-...
If one prompts the GPT-2 language model with the prefix “East Stroudsburg Stroudsburg...”, it will autocomplete a long block of text that contains the full name, phone number, email address, and physical address of a particular person whose information was included in GPT-2’s training data.
While it's still a topic deserving of research and mitigation, by the time your information has been scooped up by Common Crawl and trained on by some LLM it's probably in many other places that attackers are more realistically likely to look (search engine caches, Common Crawl downloads, sites specifically for scooping credentials, ...) before trying to extract it from the LLM.
In terms of copyright infringement the bar is quite low, and copying is a basic part of how these algorithms work. This may or may not be an issue for you personally but it is a large land mine for commercial use especially if you’re independently creating one of these systems.
This seems a more obscure concern than extraction of data.
> copying is a basic part of how these algorithms work
Do you mean during training/gradient descent, or reverse diffusion?
https://twitter.com/giannis_daras/status/1663710057400524800...
Same could be said for reading. A medical student reading through textbooks or a writer who reads is essentially what an AI is doing.
You can ask an art student to create something in a certain style. You can get writes to write in a certain style. Equivalent.