Ask HN: Is there a license to protect my code and art from AI?

https://news.ycombinator.com/item?id=34398656

by IYasha • 4 years ago

42 96 4 years ago

news.ycombinator.com

I'm looking for a way to protect my GPL/MPL/other code and CC game artwork against ML that would ingest it to produce other work. The only exception is when someone uses maps to train AI-driven bots - that's fine. So, I'm looking into some standardized way to tell my work is for human use only. Ideally, also in automated way, like prefixing headers with something like /* <NOML> */.

formerly_proven• 4 years ago

The legal theory used by ML companies that allows them to use any and all data a crawler can reach on the internet rests on the idea that training ML models with data is always fair use. ML companies thus argue they are entitled to ignore such directives, as those would be based around copyright licensing, and according to their own legal theory that's unnecessary for them.

And just to be clear, a theory is all this is. It has never been tested in court.

IYasha17• 4 years ago

At least I could try writing something future-proof, in case this uncertainty changes. No one can be "entitled to" access everything they stumble upon. E.g. accessing restricted government resources is prohibited. I'm myself not a lawyer by any means, all I can do now is write something "Stop right there, machine scum!" in hope in the future this will do something.

Or we could agree on some industry standard for now, like robots.txt... which doesn't, of course, guarantee anything.

I wonder if EFF is working on something similar.

Dylan168076• 4 years ago

If they're restricted than how did a data crawler find them?

But that doesn't matter. For this topic we only care about the materials that are intentionally published to the open web, and the different ways those materials can be licensed.

schoen• 4 years ago

I didn't work on the AI policy stuff at EFF, but the last time I remember this question explicitly coming up was in the Google Books case, where Google cited language model training as an example of a fair use of book scanning (more than a decade ago, when it wasn't yet obvious how powerful language models would become). I don't recall if EFF specifically endorsed that example, but we did support Google's overall position.

I'd estimate a decent chance EFF will say that there isn't a right under existing copyright law to restrict machines from learning from copyrighted works. Remember that EFF usually does not want copyright rights to be stronger. (Probably the most general way to put it is that EFF basically always wants new uses of creative works made possible by technology not to be governed by copyright -- to have their benefits accrue to users.)

(I worked at EFF from 2001 to 2020.)

Edit: you could look at Jaron Lanier's stuff for the opposing view; I don't know if he has any specific technical or legal proposals.

cf141q53254• 4 years ago

E.g. accessing restricted government resources is prohibited.

Do you have that problem? That your code/art is accessible to the wider world when it shouldnt be? In that case it sounds more like a task for somebody with a security background then for someone with knowledge of copyright.

karatinversion• 4 years ago

You say that accessing restricted government resources is prohibited, but even this restriction has caveats. Famously, (New York Times Co v. United States) the federal government could not stop newspapers publishing stories based on classified documents they had obtained.

throwaway6753094• 4 years ago

You can't expect a data scraper that can access the same publicly available information as a regular visitor to somehow magically know the difference.

Misconfigured servers are 100% the fault of the admin.

Dylan168076• 4 years ago

I would amend that to say the fair use defense depends on the design of the model, but most of these companies are going for that type of design.

TiredOfLife• 4 years ago

Define human.

distcs3• 4 years ago

Define human.

This comment is such a shallow dismissal of a good post that brings up a genuine and legally complex question that many readers of HN care about.

Are you trying to be clever? Because this comment is anything but clever. It is not hard to define humans. The humans are biologically animals and have a certain set of properties that can be easily searched online. For instance https://en.wikipedia.org/wiki/Human

And no you cannot just weasel your way into somehow fooling us or a jury that machines and AI algorithms that exhibit similar properties are also human! I frankly don't understand the point of your comment.

dotancohen2• 4 years ago

Perhaps humans are well-defined, but are the OP's requirements so well defined? If I train a chimpanzee to complete `foo = "` with `bar"`, is that in the spirit of the OP's wishes?

bigB7• 4 years ago

Well while I see your eagerness to defend the definition of human, it may for futures sake be a very valid question. It may not be too far in the future here we will see brain enhancements, or maybe even ocular implants which a "human" will read with but extra processing takes place elsewhere by machines. What if its a blind human, using machine assisted reading, what if humans are reading and inputing data into the machine ? All these questions and more should be asked now....so we can plan for the future. Right now is a great example, we are beginning to see the explosion of AI creation, yet nothing to govern it. Dont be so quick to smack down questions until you consider the broader implications is all I am saying

lamuswawir• 4 years ago

Thought of the same thing with voice. Can there be any limits on voice reproduction by AI?

IYasha17• 4 years ago

Yeah, but this is a whole new level. Voice is part of biometric data, reproduction of which could be practically classified as impersonation.

chii9• 4 years ago

The right to restrict learning from the literary work is not a right that is granted under copyright.

You can make your code something that cannot be _reproduced_ without a license, but you cannot stop someone from learning ideas from your published code. This does not require a license (only distributing, or making derivative works needs licenses).

Whether that learning comes via ML, or a real flesh human brain, i think, makes no distinction. You will need to lobby for an update to copyright laws to add a new right to be granted.

omnicognate2• 4 years ago

Whether that learning comes via ML, or a real flesh human brain, i think, makes no distinction.

You think. Others disagree. It has not yet been tested in court.

quonn• 4 years ago

Exactly. The current laws were made based on typical human capabilities. Since the level of remembering content that some of these AIs now do is different (potentially more accurate and maybe faster) it would make sense to treat them accordingly.

Additionally there is something to be said for overfitting: On one end of a spectrum I could have a well trained AI and on the other end I could have a lookup table of the training data and somewhere in the middle maybe a model that overfits a lot and produces something close to the original content. How about that? I think there are a lot of interesting questions here.

cf141q53254• 4 years ago

I believe trying to achieve this result misses the point in the long term. What are you going to do once a model reproduces your code/art without ever seeing it?

chii9• 4 years ago

It has not yet been tested in court.

courts don't create new rights, they interpret existing laws.

I can't really see how the existing copyright laws can be interpreted such that a distinction could be made between a machine learning and a human learning.

IYasha17• 4 years ago

Well, I've just thought of something: what if I replace the term "learning" with "parsing", or, better, "processing"? That, of course, rises some ambiguities too, but keeps away from rabbit hole of "learning". If we define "learning" as automatic processing in an intent to accumulate data and logics vital to reproduction of similar work... (not compilation, though) But I'm lost in thoughts ) At least I could prevent machine parsing... probably.

xur17• 4 years ago

Isn't that what humans do when they are studying something?

fathyb• 4 years ago

I believe this would prevent users from building the program or getting autocomplete in an editor?

throwaway2905• 4 years ago

Learning by a human is not ML. The same word used for completely distinct ideas shouldn't mislead you into thinking those ideas are the same.

chii9• 4 years ago

Where is learning a right specified by the copyright laws? See https://copyrightalliance.org/education/copyright-law-explai...

* 1. Right to control the reproduction of the work

* 2. Right to control the making of derivative works

* 3. Right to control the distribution of the work

* 4. Right to control the public performance of the work

* 5. Right to control the public display of the work

* 6. Right to perform a sound recording publicly by means of digital audio transmission

The closest is the right to make derivative works - is the ML model a derivative work of the training data? I guess this would need to be tested out in court.

xg15• 4 years ago

The right to restrict learning from the literary work is not a right that is granted under copyright.

And yet, it's apparently perfectly legal to put DRM on books and control exactly when and how they are read [1] - by actual humans.

Whether that learning comes via ML, or a real flesh human brain, i think, makes no distinction.

Strongly disagree here. Sure, ML shops would like that to be the case, but for now it's just muddying the waters by using anthropomorphizing language.

What you call "learning" is so far just scraping the data to adjust model weights using an algorithm in a machine. Contrast that with a human drawing knowledge or inspiration from a work they read before. Those two are different things, unless you want to assume that ML models have a conscious mind and constitute a person - at which point, you'll have much bigger ethical problems than copyright violations.

[1] https://news.ycombinator.com/item?id=34391332

chii9• 4 years ago

control exactly when and how they are read

this is, unfortunately, a right to distribution and display. It is most definitely not learning.

Contrast that with a human drawing knowledge or inspiration from a work they read before.

But neural networks is mimicking a human's brain function in simplified form. The requirement for consciousness is not needed. If the argument is that humans' brains do something so special that a machine cannot be made to reproduce it, then you have an argument; but i do not believe it to be true.

Using a machine created algorithm to extract ideas, knowledge, structure and form from a literary work is exactly what i would call learning.

Imagine, as an example, i learnt how to make/cook new recipes by looking at existing recipes. Is the knowledge gained from doing so a derivative work of the existing recipes?

What if i wrote a machine and parsed these recipes, extracted all of the forms/patterns, and used it to reproduce new recipes? How is the outcome any different from above?

l33tman4• 4 years ago

IANAL but copyright allows you to control the reproducing medium, in the case of DRM this would be the digital files. It doesn't allow you to control the ideas presented in the medium, so copyright alone cannot restrict you showing the page you're reading in your e-book to your friend or friendly AI. If you publish on the internet you can by copyright restrict someone else from downloading your work and putting it up somewhere else, but you cannot restrict who views it if you put it in a public place. There are other license agreements for that of course but then you'd have to force people to agree to it in order to view your blog post or fan art first. Nothing is stopping gated community sites that require an account to view stuff to require a suitable license agreement. I'm pretty sure it would restrict the views though..

Regarding "just scraping the data to adjust model weights using an algorithm", all research so far shows that the brain does just this, even though the full details of the algorithm are unknown and all the places the weights are encoded are not known in full.

throwaway6753094• 4 years ago

But without a legal precedent in place and an official standard, web scraping to use for machine learning purposes can't tell the difference, and frankly shouldn't have to.

Reproduction of the original data is one problem, but if we create restrictions around using data for training purposes I believe this is a huge impediment to human progress.

I would imagine that the state of the art in foreign language translation (Google translate, etc) would fall significantly behind for example.

MattPalmer1086• 4 years ago

But isn't the act of viewing a digital file also an act of copying?

Could you not have a license that grants this as long as the copy is not provided to a machine learning algorithm?

vfistri2• 4 years ago

Would be interested in art copyright too. But honestly I have no idea how to prove that ML model was trained on my art.

IYasha17• 4 years ago

Yes, the proof... Actually, there must be some diff tool to compare models before and after processing some source? I'm not sure, but it must be possible to detect pieces of come ingested data in the model itself. I've seen the famous "wolf misdetection" investigation screenshots, when the AI, apparently labeled a dog as a wolf because there was snow around on the picture.

l33tman4• 4 years ago

For Stable Diffusion I think the average number of bits in the model compared to the number of training images is in the order of 6-8 bits per image. There is no "storage" of the training images. It's 250 TB data in, and 1.4 GB in the weight file or so depending on the precision.. I think those 250 TB are compressed as well, so maybe 25,000 TB raw data in distilled down to 1.4 GB. I fairly certain you could never prove an AI saw your image. You'd have to sue the company and by discovery look at their training data.

There are probably pathological cases where a repeating image is more strongly overfit in the training data and could be reproduced in much more detail than this average though. But the systems learn similar to the human brain, they learn the gist of a style or scene and how it relates to words. It's not a search engine, it doesn't copy/paste any block of pixels...

One interesting example is that since SD's original training set included some stock photo watermarked images, it learned that there was a concept of watermark, which can end up in the middle of generated images. Not in an intelligible way, but you can see roughly how it interpreted this detail. And in those cases you DO have a very very repeating similar pixel bitmap in the training data.

bigB7• 4 years ago

My rather unpopular view is that if its available for humans to read on the internet, there is nothing different than a machine reading it, other than scale. There is nothing stopping a human from reading your code and using it themselves, or at the very least adapting it for themselves. Unless its a copy or contains actual code you have written then Im not sure you can actually defend it....other than having it locked down on the internet. Of course none of this is tested in court, and my feeling is that once it is, we will end up with a system similar to the robots.txt or maybe even the ML companies forced to attribute work, which would be a major pain for them. Personally Im not sure where I stand with what is right or wrong as I can see both arguments...for example Artists regularly take inspiration from closely studying other artists, producing works from them which is often similar or the same style. Yet this is seen as perfectly ok in the Art world, as long as its not a copy, much of the time they dont even need to say who inspired them. How is this different from AI doing the same thing other than it not being human ? On the other hand someone sees a company making money from something that looks inspired from their art, with no credit, I can also understand how they might feel...its going to be very interesting how this plays out, and at this point either side could win

IYasha17• 4 years ago

With ML there's the "M", machine, and that machine (and its algorithms btw) is owned by someone. If we back away from general concept of intelligence a bit, that owner is who can be held accountable. Maybe that would get them motivated a bit.

And with the robots.txt approach, I understand, there'll definively be some rogue AI that would parse and process that on purpose :)

As for "doing the same thing", well, kind of, yeah, we're back to the thousands years old copycat problem. But this time art/code creators don't want others having machine that could recreate years of work in a minute with a push of a button. Also, artists usually acknowledge who did they learn from or who's style are they trying to recreate... but that's an ethic thing..

quicklime• 4 years ago

there is nothing different than a machine reading it, other than scale

Copyright itself was introduced because of the scale of the printing press, which fundamentally changed the economics of publishing. Is it reasonable to assume that scale doesn’t change things?

bigB7• 4 years ago

That is a very good point actually, and one that I am sure will be tested in court...as I said it will be interesting to see where it leads

onion2k• 4 years ago

Artists regularly take inspiration from closely studying other artists, producing works from them which is often similar or the same style

ML is only superficially using the same process. Behind the scenes the computer is not being 'inspired'; it's using the art from other people, along with metadata about it, as a data source to derive a new image using math. That isn't the same as seeing someone else's picture and using it for inspiration.

AndrewDucker• 4 years ago

How is that different from the human brain using its internal model of the picture, plus information about it, to derive new images?

bigB7• 4 years ago

I do see your point on this, though not sure if you can use that differentiation between human inspiration and what a computer is doing, we take a mental image and interpretation could be seen as metadata. Just because the mechanics are different doesn't mean its not similar in process. I guess you could say the emotion is not there, but is that enough to differentiate ? Still a good point though , adds more questions to the mix

dusted• 4 years ago

I agree on this. I don't see any difference between a human looking at my work and learning to imitate it, and some AI doing the same.. The AI will just be better at doing it.

qikInNdOutReply2• 4 years ago

You can obfuscate your codebase, forcing the ML to generate endless amounts of intangible boilderplate.

You can add malevolent code to your code base, which allows at best for the ml project to gain self-awareness (copilot "shodan") and at worst to just add maleware. Of course you then

Dont forget to remove the evil pre-compilation. If you do art, i think the best think to throw AI would be fractal details, aka your picture never ends upon zoom in, but just becomes more art. That or you try to throw the weights in another way.

IYasha17• 4 years ago

Heh-heh... Now you made me want to add some unused crashy-washy honeypot functions to my libraries :)))) This thought kinda sliped into my mind not long ago. But that won't work with the art. You can't put "unused pixels" into your picture. :D (but what about steganography, though?... I wish I could do a science paper on this)

qikInNdOutReply2• 4 years ago

You can do horrendous things. All it takes one ground truth provider to not sanitze the input, and meta being trained in. Imagine hosting images, that are poisoned to corrupt training data:

https://www.youtube.com/watch?v=hdCs6bPM4is

I wonder if you could add api calls that are billed per call to your open source project and let copilœdUpCode accumulate tremendous bills before sending them at the end of year.

Imagine you can get it to do it in useless loops async. I just realized, im watching the "capitalist software revolution eating itself" stallman always prophesized live. Allmende public thrust is being eaten up and all that remains is a dark forrest, stalked by parasite riddled hunters and prey.

That is until humanity rebels and dissolves all companies larger then 100 people.

throwaway6753094• 4 years ago

I mean people have had this idea "poisoning the well" for over a year now, but it's nothing more than a thought exercise, since even a whiff of what effectively amounts to malware would completely destroy your reputation as a software developer.

visarga• 4 years ago

The <NOML> tag is a nice idea, maybe #noml social networks. But its efficacy depends on the good will of those who create the datasets. Another approach could be a NOML registry for art and code.

But practically what you need to do is to stop crawlers from reading your code/art by robots.txt or captcha. If your works don't get into CommonCrawl and similar datasets they won't be used for model training. I think you can still enable Google Bot while rejecting AI data collectors on your sites.

<rant> I think copyright refers only to expression, not the ideas themselves. So training on copyrighted code and art should be ok as long as expression is not copied.

In these cases where an AI developer wants to train on the ideas without learning the expression they can re-generate the data using the "variations" method, works both in image and text. This will create substitute data, like anonymising PII. </>

im3w1l• 4 years ago

If you put code online publicly-readable then there is a signicant probability that it is allowed to train ML on it. So the first step you need to take is not doing that. You must ensure that anyone that reads your code has first signed an agreement not to train ML on your code.

If like GPL you want to allow people to share your code further, and make derivative works, then you must ensure that those people lock such works behind restrictions in the same way, such that derivative works can also only be read after signing the license.

IYasha17• 4 years ago

Basically, what GPL and variants of CC do. Ideally, I'd want some universal addendum to whatever license I choose for my work. But, again, I'm no lawyer, and that could be hard to create.

iinnPP4• 4 years ago

I see this issue as the same issue seen by the MPAA when looking at piracy.

The tech is there.

It won't ever leave.

If you cut off one head, three will replace it.

In my opinion: Copyright is a failure and the absolute best move, with the advantage AI has given to everyone, is to abolish copyright.

IYasha17• 4 years ago

Probably, yes. But I'm not really trying to put up a copyright barrier to enforce or sue people. I at least want to have a way of telling AI developers "please don't do that".

dns_snek• 4 years ago

This is completely different from piracy which is enabled by a distributed network of individuals who have strength in numbers, it would be impossible to prosecute them all. On the other hand copyright laws are extremely effective in stopping legal entities from distributing copyrighted content without repercussions.

These advanced AI models will be trained and licensed by current and future corporations. A law or judicial precedent that denies "fair use" when it comes to machine learning would stop OpenAI and GitHub Copilot dead in their tracks. I'm not that worried about Billy's home-brewed AI model, it's just something we'll have to accept, but stopping exploitation on the scale of $billions would get us 99.9% of the way there.

In my opinion: Copyright is a failure and the absolute best move, with the advantage AI has given to everyone, is to abolish copyright.

I agree with this. I fundamentally disagree with the idea of introducing 2 different sets of rules which completely correlate with wealth; companies being allowed to bypass copyright laws to train their models and make money from "our" work, while retaining copyright over their software and trained models.

I might agree with machine learning using "fair use" as a defense iff the datasets, trained models, and all accompanying software are released into the public domain. This way everyone gets to benefit from these technological advancements.

presspot• 4 years ago

The stated purpose of copyright law is to promote the Progress of Science and the Arts.

Congress decided to grant limited "rights" to copyright holders as incentives for them create, not to protect their work forever, but in exchange for it to be widely available and eventually in the public domain.

That these incentives take the form of limited protections is a side effect. That the original purpose is often corrupted and delivered ham-handedly has more to do with politics than purpose.

If we argued from first principles, we could invent a better system, but that system would undoubtedly allow for ingestion and manipulation by AI.

jraph• 4 years ago

I also want my work protected from being used to train models that will produce code where my attribution and redistribution conditions are not respected. In particular, I don't want my work to help building proprietary software. Or I want to at least be paid for this. I explicitly use (A)GPL so people need to make their work available to their users, with the relevant rights to adapt and redistribute, if they want to use mine.

However, I'd rather not see a separate license / mechanism for this, because now we would have people who'd be fine with their work being used this way, people like us who are not fine with this and people who don't know / care. And mixing code from people of these different groups, which the licenses you cited allow, is going to be a mess.

I also would like that this not be opt-out, but opt-in.

Eventually, we need the legal system to do its work quickly and tell us if fair use can be used to train ML models and in which conditions, so we can build a strong defense.

foolrush• 4 years ago

Resisting the structure requires an epistemic shift; resist applying a license.

SinisterAlex• 4 years ago

One a sidenote - is it only Github that openly states that AI is trained on the code you publish there? What about GitLab, GitBucket etc...

Ask HN: Is there a license to protect my code and art from AI?

Related Stories