I don't know the name, but I remember some sci-fi story about some academy where humans were trained from birth without exposure to music others had written and had to reinvent it on their own. Some would cheat and access the outside world's music, but they would always be caught by their later compositions all having obvious influence from conventional music.

<Insert obvious joke about how I'd remember the name if I had computer-like memory.>

You can read it here: https://b-ok.cc/book/4395497/b2fb2e

It turns out that humans can extrapolate generalisms to a degree we are currently unable to explain clearly enough as a model to imitate.

It turns out that much ML is merely referenced regurgitation.

Marketing and hype are rather advanced skills in 2022, however…

But this can't be right, it is inconsistent with how copyright has worked so far. Artists and musicians and engineers all learn from each other and have seen and learned from, "trained on" many other examples of works from their field. Even when works are clearly inspired by other works we tend not to give them the legal status of derivative work.

You're suggesting we treat models with a much stricter copyright regime than has previously existed.

If we get to the point that capitalist maximalist-utilitarianism insists upon hijacking the very concept of what a living organism is, I can only compare it to a teddy bear vs an actual bear.

It’s not enough to merely put fur on it and an internal rom for it to regurgitate prefabricated roars upon contextual prodding.

Respect the life you are only one instance of, for hubris has always brought suffering and pain in its wake.

You have a mechanism that can regurgitate (digest, mix, emit) without attribution all of the code and all of the art.

With these systems, you're giving everyone the ability to plagiarize everything effortlessly and unknowingly.

Everyone can use my code, without ever knowing I wrote it, without ever knowing I existed. Everyone can use my hard work, regurgitated anonymously, stripped of all credit, stripped of all attribution, stripped of all identity and ancestry and citation.

It's a new kind of use not known (or imagined?) when the copyright laws were written.

Training must be opt in, not opt out.

Every artist, every creative individual, must EXPLICITLY OPT IN to having their hard work regurgitated anonymously by Copilot or Dall-E or whatever.

If you want to donate your code or your painting or your music so it can easily be "written" or "painted", in whole or in part, by everyone else, without attribution, then go ahead and opt in.

But if an author or artist does not EXPLICITLY OPT IN, you can't use their creative work to train these systems.

All these code/art washing systems, that absorb and mix and regurgitate the hard work of creative people must be strictly opt in.

That's how the law needs to be.

This would prevent anybody from accidentally infringing when using these tools. Does that seem like a reasonable solution, or is your concern greater than accidental infringement?

A world without copyright would however save more than just a few millions of man hours that copilot might do. Allowing people and companies to freely use the best software available, view the best art, enjoy the most relaxing music, have the best recreational time with the best films. The only harm is the highly speculative claim that people won't be creating the best software, the best art, the best music or the best films.

In those cases it seems that humans are already copying code without also propagating licenses appropriately. LLMs are more likely to memorize things which occur a lot (and I'd bet rare things that are representative of some conceptual axis).

The main examples presented so far, Davis and Carmack, have the property of having been copied a lot. The generative model is only surfacing an existing pattern of ignoring attribution. Sort of like the code-gen version of generating bigotry if appropriately prompted.

I'll also note that this pattern of retrievable memorizing of copyrighted and sensitive material is present in GPT-3 too and not just for code. As the situation is equivalent, a lawsuit should address the concerns of non-programmers too.

An open equivalent might be wary of being accused of contributing to copyright violations since in that scenario, there is no way to force people to respect it.

Kinda the same problem as YouTube. Lots of people copy movies on the high seas, but if you are as big as yt you cannot easily get away with it.

This is not like youtube because Github is already hosting those violations and people are already inappropriately copying or including such code. It matters not whether the local inclusion was fetched by copilot or a human fetched it using more manual steps through search.

But in defense of copilot, code regurgitation is uncommon in routine use. An editor extension allowing search of github would be at least as easy to use to violate licenses but I do not think it'd be taken down since that would not be its core offering.

Copilot goes far beyond mere search and provides a useful service. GPT-3 can also be prompted into generating copyrighted works of writers but I do not see people talking as if that is its primary utility nor as much clamoring in these forums to end that service.

Well, try doing that for music or movies or proprietary leaked codebase.

If you think copilot is uniquely producing things that are not that different from humans then surely no one would have any problem with feeding it massive amounts of corporate programs?

I am not aware what writers are doing but there have been plenty of uproar regarding stable diffusion. I have a feeling that if any tools like this get built for musicians/film-makers, it will look vastly different from the current situation.

> surely no one would have any problem with feeding it massive amounts of corporate programs?

There is a similar gymnastics done by human engineers today due to the issue of patents. I don't think this is a good trend to uphold.

> I am not aware what writers are doing but there have been plenty of uproar regarding stable diffusion

Yes but mostly in the art community. On HN there were plenty of arguments just the other day how it is not the same for art and programmers have a stronger case. I disagree but regardless, the case is exactly equivalent for GPT-3 and writers but it wasn't an issue generating round about a thousand comments on respecting IP until copilot.

Are you arguing that copyright laws should be abolished? I have no problem with that as long as it's clearly defined, you can't not respect copyright of open source code but enforce it for proprietary code, just because the value is arguably non-monetary.

This is quite important, actually, and I don't think enough people realize this. I am a photographer sometimes and it would be really cool if I could share my photos online under a copyright license that forbids their use in training AI.

Or they could integrate a way to find the produced output back in the corpus if it's sufficiently close and provide a reference/attribution. Basically whatever tool a copyright lawyer would use to track down original work.

And that's just the engineering solution. The AI researcher solution would be to extend AI learning algorithms to attach attribution metadata to the learned data so that the output could already come annotated with information about the source.

But the latter is much harder to do, so maybe the engineering solution would suffice.

Edit: here's the related tweet: https://twitter.com/DocSparse/status/1581632706693079042

That assumes that the licenses of your code and the original code are compatible which often isn't the case.

A user once replied to one of my comment[0] about this with the following:

> It's not really an issue when you're a large software corporation; you already have mechanisms in place to check for license compliance in everything that ships, including F/OSS plagiarism checks [1].

IOW, from my understanding, they don't care. Big players do their own checks anyway, and small fish won't be creating problems because it's too convenient for them. Classic Microsoft (as I know from 90s).

The bigger thread can be seen in [2].

[0]: https://news.ycombinator.com/item?id=32534697

[1]: https://news.ycombinator.com/item?id=32539467

[2]: https://news.ycombinator.com/item?id=32533531

Copyright (unlike patents in general) allows for independent creation. If I sit down to write a quicksort routine, it is going to look extremely similar to a zillion other quicksort routines out there.

The other question (IANAL) is whether writing a quicksort routine is even a creative act at this point.

Which amounts to saying, humans are trained with a model that they can use to recognize when something they are thinking of producing is 'too similar' to something they have seen before.

And, of course, some humans choose not to apply that filter and go ahead and plagiarize anyway; some humans try to apply that model but get back a false negative, thinking they're producing something original when they aren't. And we have ways of dealing with humans who do that.

In the case where an AI is coming up with the work, perhaps the mistake is in relying on humans to try and apply their own trained judgement to figuring out if the result is unoriginal. We need an AI that scores work for how likely it is to be infringing on a prior copyright.

Then you use that AI to train the creator AI, and teach it 'originality'.

> In the case where an AI is coming up with the work, perhaps the mistake is in relying on humans to try and apply their own trained judgement to figuring out if the result is unoriginal. We need an AI that scores work for how likely it is to be infringing on a prior copyright.

Isn't that latter AI going to be more likely to need to contain verbatim copies of original works? Or maybe not?

This in turn (and the SFC's and the law firm's concern about the GPL) makes me think that there are several different things that people may be concerned about machine learning systems doing:

* they could allow you to access verbatim copies for "consumptive" use (like if you asked an AI a question about what the text of a chapter of a Harry Potter novel was, and it answered you correctly)

* they could facilitate intentional or unintentional plagiarism, and, in the case of publicly-available works that are published under a license, intentional or unintentional reproduction or creation of derivative works contrary to that license

* they could contain something like a verbatim representation and allow you to use that in various ways that themselves are not extracting or literally copying that representation, but where the original copyright holder might complain that the existence of an unlicensed copy inside the model is already objectionable

* they could contain representations of uncopyrightable subject matter which was learned through training on copyrighted works, which can then be used to compete with the original creators for jobs, prestige, or attention, or can be used to produce works that the original creators would have found offensive or objectionable (this case isn't supposed to be restricted by copyright at all, but that doesn't necessarily stop people from caring!)

Not only will the same measures not prevent or avoid these cases, but if you wanted to prevent the first and second situations, one of the easiest ways to do it might be to literally include verbatim copies of lots of works inside a machine learning model! (along with software specifically trained or programmed to warn you against unintentionally making uses the user or copyright holder finds objectionable ... to facilitate the exercise of "essentially, their ethical judgement", as you put it)

I'd be wary of this one. There doesn't seem much distance between this and the same claim against a human's memory of a work in their own brain. Yes, that sounds like dystopian fiction. So do some things that have already happened.

The big question is if we think that Google made a poor work of that AI and if more money and more data rich company can make a better AI that teach originality.

> It seems clear enough to me that training AIs on copyrighted works is typically or commonly a fair use under existing law, because the AIs can and commonly do learn non-copyrightable elements and aspects of those works. It's very obvious from enormous numbers of examples that current AI systems are capable of learning much more abstract features of human culture (grammar, concepts, facts, cultural tropes, and many others).

I said it already in a previous discussion, I would be very careful with comparing ML with how humans learn. To me there are still a lot of examples that show that AIs don't understand prompts (see e.g. the discussions around the "horse riding astronaut" prompts för stable diffusion et al.) and it seems like they really are just doing sophisticated pattern matching. If that is what they do aren't they themselves covered by the licenses/ restrictions placed on the "patterns" they "choose" from?

Capitalism is still very alive, and will continue to be. It's in conflict with the general welfare of the people...

Something will have to change.

I mean sure, maybe humans aren't "just" doing sophisticated pattern matching, but there are good reasons to suspect this is some part of what we are doing. (Even if its not implemented with back-prop).

e.g) consider the work of people like Anil Seth, who propose that our brain is basically a generative model of the world, which aims to minimize the likelihood of perceptual data. (see also: Karl Friston's free energy principle). What's up for debate is how it is structured, what priors are built in, what is the learning algorithm etc.

Anyway, for all their limitations, it seems clear that current artificial generative models can: 1. learn hierarchies of abstractions, which 2. explain the observed data in the fewest possible number of bits, and 3. generate new, novel data based on the patterns that have been learned

If you want to describe this as "just sophisticated pattern matching".. then sure I guess? But I think there's a clear qualitative difference between this and searching for code in a discrete database (which imo would not be okay).

The paper suggests that transformer work on a set of select, aggregate and element-wise operations. Which seems pretty close to the SQL statements i write from day to day.

And at this point, you are getting both attacked and supported by AI, and not really better off in a meaningful way.

There's no good way of solving this issue, without general intelligence, and the problems that will bring (what reason does a generally intelligent AI have for supporting us or not enslaving us).

This is why all AI research IMO is unethical. Point me to a single AI use that has not already been abused, maybe I will change my mind, as it stands though we should be prosecuting the people misusing this technology, or at least irresponsibly releasing it, as fast and as quickly as possible before we get the point we are no longer fighting bad actors but the machines themselves.

I don’t think AI research in a vacuum is as deeply unethical as you suggest. It’s about current societal context- people won’t like being worse at things; resources will be hoarded rather than distributed.

I think OP explains clearly, in many paragraphs, why it's fair use. That's literally what their whole post is about.

Clearly, training on large volumes of data is not small volumes in any sense of the word. The argument that it is fair use is itself flawed.

Or here is the real analagous question:

Fair use is about more than just the size of the excerpt.

If you write an article about good writing, and quote a choice paragraph from someone else's work to show an example, and credit that quote, that is fair use.

Is it fair use if you read an awesome paragraph, something that really is the result of the authors unique intellect and effort and craftsmanship, and makes you think "damn", and then drop that same jewel into your book?

You can probably get away with it, because you probably just won't be able to convince a judge that any single paragraph is that big of a theft.

But I don't mean to ask if you can get away with it, I mean to ask if it should be considered fine honorable behavior.

The difference is, the paragraph isn't being included for examination or comment or transformation, it's being included to directly copy and perform it's original function as part of what makes a work a great work, and, it's not being credited in any bibliography.

This is what copilot is doing, or rather it's facilitating people doing it, as far as I can tell.

The original snippets are functional not there for examination, copied verbatim not transformed (sometimes), and not credited.

Most of it comes from open source works anyway and most authors would probably be fine with it if the stuff was simply credited.

I think in programming therms a useful parallel might be copying at the module rather than the statement or function level. For example, if I write some code prompts to do the following:

  - validate my API key with Twitter
  - solicit the input of a Twitter username
  - download the up to 500 of that user's tweets
  - convert the json to a dataframe
  - plot the derivative of the intervals between tweets

Actually, what the OP said is, "is typically or commonly a fair use under existing law, because the AIs can and commonly do learn non-copyrightable elements and aspects of those works". The rest of the eight paragraphs had nothing to do with fair use.

It's honestly a ridiculous argument to say that learning one non-copyrighted thing means that the regurgitation of another copyrighted thing, after stripping the license, will magically be fair use.

FTA: "On the other hand, maybe you’re a fan of Copi­lot who thinks that AI is the future and I’m just yelling at clouds. First, the objec­tion here is not to AI-assisted cod­ing tools gen­er­ally, but to Microsoft’s spe­cific choices with Copi­lot. We can eas­ily imag­ine a ver­sion of Copi­lot that’s friend­lier to open-source devel­op­ers—for instance, where par­tic­i­pa­tion is vol­un­tary, or where coders are paid to con­tribute to the train­ing cor­pus."

This is the same argument that people use about Stable Diffusion, and it's kinda meh to me...I guess it'd be nice to allow people to opt-out, like Stable Diffusion is doing with their next versions, especially since a negligible percentage of people will do so and it won't affect the models at all. But yes, it basically is yelling at clouds. Opt-in would cripple models, and some people would make them anyways and just keep them secret, which is worse for the world. And at the end of the day, this really does just seem to me like a fair use of stuff that you've published on the Internet for anyone with a browser to look at. The AI models of the future are going to gobble the whole net up, and if you don't want them ingesting your stuff and learning from it, then you just shouldn't make it freely available.

If OpenAI/GitHub/MS really wanted to get ahead of this and head off any potential legal conflict, they could always just open source the models and weights, which would be in line with the name "OpenAI"...it would be a minor project to scrape all the correct headers to add to a license file(s), but negligible compared to the many millions of dollars spent on training.

Also, it's pointless to say "But X does Y" in copyright discussions. You never know if they license the content properly or if they infringe the rights. In the Cliffs Notes case, they might not need fair use at all, because the old works are already in public domain.

If the AI is learning to repeat text (e.g. Copilot) or images (e.g. Dall-E), then that makes it possible to reproduce the copyrighted works, so I would agree that that case is not fair use. -- It would be akin to compressing and distributing those works.

If the AI is learning patterns -- such as "muggle" being a noun that relates to Harry Potter, or that the lemma for "muggles" is "muggle" -- then that is less clear. You can avoid the situation by creating your own sentences with those terms in them, and annotating those sentences instead of the copyrighted ones. That way, the AI is still learning the same information.

Because copilots "use" of the works _was_ the learning.

So it would seem to me that Microsoft needs to apply "fair use" to copy and redistribute _the entire works_ they used for training.

In which case lack of fair use my well be the least of their problems, they are really crossing into Computer Fraud and Abuse Act territory similar to when Aaron Swartz "borrowed" MITs data.

This issue has been claimed many times and I've heard that DALL·E 1 & 2, Stable Diffusion & Midjourney all can create images that are exact copies of the training material.

This doesn't make sense considering the compression ratio of training images to model is about 1:25,000.

Further investigations I have made show that all these cases can be explained via the following:

1) The prompt included an image, so some form of image2image was used. Of course if you use an image as a base, and tell the model to stick closely to that image, the output will largely resemble that image.

2) The example was completely made up.

So far I have seen no evidence, given a text prompt, the output of an image containing some portion of any image from the training set.

Well, I don't really think you are violating my copyright. But by focusing on parts, you go down a rabbit hole of equating an element with the whole thing. This would render all collage art illegal. Lawyers and art pundits love ruminating on the uncertain legality of collage art (because it's not a binary question, so they can churn out endless articles that boil down to 'it depends'), but this glosses over 2 important realities:

1. Nobody gets sued over collage art largely because any case is doomed to end up with lawyers measuring the size of collage elements with rulers and then arguing about what small percentage is too much, and uncertain exercise few law firms wish to gamble their reputation on, and

2. nobody gets sued because collage art isn't worth very much to begin with; collages aren't valued very highly because they aren't as hard to make as painting or other art forms. 'Appropriation artists' like Richard Prince get rich and famous partly because their art is less about the image than the cultivation of notoriety for artistic effect; they are artists of scandal rather than pictures.

In general, bits of things are just not that important, and I'd argue that the same applies to code. If part of your code matches a prompt (excluding highly specific prompts like '# insert Woodson's unique XYZ algorithm here') and is then deployed in another program without alteration, isn't that most likely to be because it performs some generic function?

The first assumption is highly flawed though. Humans routinely do violate copyright law. Plagiarism is a huge problem in many sectors; un-cited direct copies of people's work in violation of fair use is a regular every day occurrence in the human world. It doesn't matter if you memorized the source material or if you transcribed it or if you copy and pasted it, if it isn't your source material and is someone else's, you've committed a violation of the law. Learning to produce original work and reproducing someone else's work is not the same thing. If an AI is ingesting and perfectly reproducing someone else's copyrighted works, it is in violation of copyright law in the same way a human would be if they reproduce someone else's copyrighted works.

If a corporation was to directly publish some copy that appears plagiarized, we'd call that plagiarism. I don't see how adding a piece of code—one that's fully created, owned, and wielded by the corporation—as an intermediary changes anything. If anything, it looks like plagiarism-as-a-service, which seems worse (at least to my eyes).

Of course, this matter is a bit confusing. Because, for example, (1) it's not always plagiarism, (2) defining what exactly is plagiarism even in the purely non-technological realm is difficult (and likely somewhat subjective), and (3) there is a lot of corporate marketing which suggests this "AI" is "autonomous" (presumably to distract from who exactly is autonomous in this picture). And of course ML art is quite useful for many things. But I mean, so are artists.

Not long ago, a lot of Silicon Valley rhetoric was that the purpose of "technology" was to free up time so that people could be more incentivized to "do what people love to do" like, for example, artistic creation. But now it seems that rhetoric was just that: rhetoric, or what was needed to be believed/said at the time.

And now at our present time, when technological "progress" has been followed a bit further (that is, when we've developed our machinery a bit further under the incentives of our present economic system), much rhetoric has conveniently shifted to something else, something largely contradictory, but again precisely to what is needed to be believed/said to continue following the same incentive structure.

Copilot is capable of going beyond retrieval and is competent at using variables, comments, types and local context to infer intention and generate appropriate code and even comment on it. Whenever copilot correctly predicts code of yours that's a novel combination of concepts, copilot has originated novel code.

For esoteric concepts, you usually already have to know how to prime it but Copilot is especially useful when it helps you bump into things you didn't know you didn't know (one way to increase the odds of this happening is to write out your thinking so far in markdown or comments. You'd be surprised how helpful and clever Copilot can be in some instances). My point here is Github isn't charging $10/month for run of the mill retrieval. My opinion is code-gen LLMs contribute value and more open versions are worth building.

That said I think the results are often useful and sometimes fascinating. We should not fool ourselves about the learning that these large neural nets do, though.

>Sure, we use the word "learn" to describe what they do, which is one word that we also use to describe what people do. But ML models are always wielded by people or corporations for particular purposes.

This is extremely important. "Learning" in machine learning is an aspirational label, not a descriptive one. People who claim otherwise either drank too much of their own Kool-Aid or are simply dishonest. This isn't just "wrong" in some taxonomical sense, this is dangerous in a very practical way. Conflating machine "learning" and human learning will inevitably lead to various kinds of sabotage of human learning.

Programmers tend to think of copyright as a Boolean valued function. Either something is infringement or it isn’t.

Judges think of copyright infringement as a real-valued function of many arguments corresponding to the circumstances of the parties (e.g. what actual damage was done?).

A human quoting a human without attribution, without any profit made or identifiable damage, returns an infringement value very close to zero. Such cases, if anyone is petty enough to bring them, are likely to get dismissed.

[0] https://en.m.wikipedia.org/wiki/De_minimis

I absolutely agree with you that copyright is not a boolean, but I don't buy the idea that a judge will just shrug and allow infringement to continue just because there was no commercial harm.

I also think your example is just irrelevant to the case at hand. Sure, someone "performing" someone else's copyrighted words once may not be a big thing. But if Copilot is actually found to be infringing, these infringements will keep happening, over and over and over.

Bottom line is that none of this has been tested in court. I think it's great that someone is working on doing just that. Maybe the end result will be that Microsoft's use is indeed fair use, and that Copilot users have no further obligations. But I'd like to hear a court decide that, not a bunch of armchair non-lawyers (myself included) on a random web forum.

(sorry i wrecked the citation and lost where i found it)

> CONCLUSION:

> The court of appeals affirmed the district court's judgment. The court held that the Performers' use of the composition, as distinct from the use of the composer's performance, was de minimis and therefore not actionable. Considering only the compositional elements, the brief and relatively simple segment of the composition used by the Performers was neither quantitatively nor qualitatively significant when viewed in relation to the composition as a whole. Thus, despite the high degree of similarity from the actual use of the recorded composition, the scope of the similarity was not sufficiently substantial to support Newton's infringement claim.

Also I'm looking for a case where the judge acknowledged that copyright infringement was occurring, but decided not to do anything about it. From the bit you quoted, it sounds to me like it's implying that the judge believed there was a valid fair-use defense? Or even stronger, that the judge just did not believe the use was infringing at all?

... as opposed to AI. At the heart of the matter lurks a debate whether AI is an independent phenomenon which behaves in its own right, or a just a tool that's created and wielded by humans against a backdrop of clear incentives and motivations.

The argument isn't about whether or not the law deals in a absolutes - it's a basic principle that law is tested in courts through interpretation - the argument is that co-pilot can be perceived as merely a means to an end and that GitHub / Microsoft have created a massive mountain of liabilities for themselves.

Suppose we add a button to a visual studio plugin called 'Copy me a function' and when you click it, it 100% grabs some random code from github and plops it as-is into your code base.

I don't have to argue the ethics of if the button is 'thinking for itself'

> Suppose we add a button to a visual studio plugin called 'Copy me a function' and when you click it, it 100% grabs some random code from github and plops it as-is into your code base.

Personally, that's exactly how I see co-pilot. To my mind, it's a tool that sits in the same category as p2p platforms, copying machines or video recorders. They are just tools.

How, for what purposes and by whom they are leveraged makes all the difference here.

P2P platforms and those who violate copyright are routinely shut down (or attempt to be shut down) in the US. If co-pilot sits in that same space, it seems the books been written already... we know how it ends.

"In computer programs, concerns for efficiency may limit the possible ways to achieve a particular function, making a particular expression necessary to achieving the idea. In this case, the expression is not protected by copyright."

https://en.wikipedia.org/wiki/Abstraction-Filtration-Compari...

Consider the impact on innovation if Microsoft or Oracle were allowed to claim a copyright over utilitarian aspects of their works such as the Java or Windows API!

BTW, Copilot seems to be reproducing copyrightable material when the tool reproduces comments verbatim!

Specifically within AFC, note the "idea/expression dichotomy" [1] which clearly states:

"copyright law protects an author's expression, but not the idea behind that expression"

Thus, if this tools spits out someone else's code verbatim it is a definite copyright infringement. If it outputs code that is similar but not verbatim then it "could" be an infringement, at your own risk, and to be determined by the courts. Simply expressing the same idea in a different way is not a definite infringement.

Program code is naturally copyright-risky because the keyword/grammar space is constrained. It is far more difficult to accidentally duplicate verbatim the expression of one's ideas in a full language, such as English, than in C. And what of two separate programs (or constituent sub-parts such as functions) that by chance emit the same compiled binary?

Personally, for now I won't use this tool due to the risk of accidental plagiarism, and because it is a black box: I can't examine any lineage or attribution metadata to understand the source(s) of what I would then be incorporating into my own body of work. Of course I doubt I could get that type of traceability information for any other trained ML model I might use, so perhaps I need to re-examine my policies heading into the future.

[1] https://en.wikipedia.org/wiki/Idea%E2%80%93expression_distin...

That is not true. It can be verbatim and not a copyright violation if it can be shown that the expression in question is strictly utilitarian! I literally provided a quote from that AFC article that says this!

There's even precedent that prior art nullifies a copyright claim, as seen in Johannsongs-Publishing, Ltd. v. Rolf Lovland:

"Johannsongs failed to offer admissible evidence to rebut Ferrara’s analysis, so there is no genuine dispute of material fact as to his conclusions that Söknuður and You Raise Me Up are not substantially similar and most of their similarities are attributable to prior art."

And this was about music, not software, which has always sat uncomfortably between utility and expression, if only because it is some kind of writing. No one is claiming copyrights over Photoshop filter settings or other inputs manipulated by sliders or buttons!

First, let's preface this with the substantial similarity of the structure, sequence and organization as established in Whelan v. Jaslow which amongst other things says that you cannot merely change the variable names if the expressive structure of the code remains the same.

Now let's imagine 10,000 software developers who all implement Dijkstra's algorithm in C and then run it through clang-format. Aside from variable names, isn't it safe to assume that many of the implementations are going to be exactly the same?

So if it's identical it might or might not be a copyright violation and even if is totally different it still might or might not be a copyright violation... and only after spending insane amounts of time and money in the court system can you ever be sure if what you (or your AI) created has made you a criminal. This is increasingly sounding like a very very broken system.

However, the statement was only for cases of verbatim copies produced by Copilot. The AFC Wikipedia article states that "Proving copyright infringement requires proving both ownership of the copyright and that copying took place." The 3 detailed tests developed in that case appear to be "expand" the determination of infringement to close potential loopholes where, while there is not a verbatim copy, infringement is still deemed to have occurred because of "substantial similarity". e.g. someone copies a program but changes the variable names.

So where is the line between infringement and not, in cases where there is an exact copy of a code fragment? Can we still use the utilitarian defense or is that only used by the court to exclude portions of the code in the tests for "substantial similarity"?

At this point Copilot is awful at this higher-order level of abstraction but I can see a time where this is not the case!

Microsoft will have to put more work into filtering out responses that are indeed copyright violations if they want people to use their tools.

I doubt that MS will ever be held liable for the violations themselves as there is precedent in themselves and their legal department has plenty of cash to burn.

And we need to accept that and get over it, not get better at outlawing it.

Likewise, you can tell Copilot to crank out code specific algorithms written by specific people, but if you do so, you're still creating infringing code, same as if you'd taken the more direct route of ctrl-c+ctrl-v. The fact that you /can/ make the algorithm misbehave through adversarial input is irrelevant to the primary use cases which lead to boring non-infringing code completions.

Your argument just disallows discussing the problem while doing absolutely nothing about it.

If you train your dog to NOT attack random passersby and it still does, that dog is euthanized no matter your intentions.

There is plenty of space here to discuss developing tools to check for unintentional infringement. I would guess, though, that such tools would sweep up a whoooole lot of non-copilot human usage and make it much harder to deploy anything new.

So, maybe a better discussion to have here is how to make the animal safer, not the total outlawing of the animal. Single-line completions (the majority of co-pilot usage) aren't infringing. Probably true for almost-any few line completion. So, capping the amount of consecutive auto-completed code might be a reasonable 'muzzle' on the model to keep it reasonably safe.

Now my software has been assimilated into a proprietary blob. Had that blob been free, like my software within it, I would have accepted it, but it's not. It's controlled exclusively by Microsoft and OpenAI, two entities which I place no trust in.

For me the dog has already bitten. The free software I extended to an audience I believe would show the same generosity has instead been made into a proprietary product.

The "copyright" question for me is not a question of "fairness" or ability of Microsoft or anyone else to make a product. For me it's a tool to protect my contribution from proprietary business.

Basically. I dont want the animal safer, I want it free (according to the FSF freedoms).

That's exactly my complaint about Copilot. And since all code hosted on GitHub is now subject to this land-grab, my only recourse is not to use GitHub any more if I want to publish a project of mine.

Your code isn't being run by Copilot, as such; it's been categorized in a way that allows partial retrieval without the license or attribution. This might seem like a distinction without a difference, but it's kind for a ship-of-theseus problem; probably nobody is running any of your programs in their entirety, but it's very possible that bits of your code have found your way into other people's programs. How do you distinguish between contributions that are uniquely yours, and those which are just helper functions or cobbled together from other example code, eg in documentation or from a book or Q&A website?

With copilot, you are not acting with knowledge of the source and not with intent to copy, in fact I'm sure the users would have a reasonable expectation of the tool not copy-pasting existing code verbatim.

IANAL, but I'm pretty sure that intent matters a lot.

Popcorn time was also just a tool to allow you to stream data from torrents. That didn't seem to help them put up a legal defence (nor should it have, because the intent was pretty clear on that one).

And seriously, if cases exist, where the only thing a tool does (albeit via a VERY complex implementation path) is to strip a license from a piece of code and serve that code up via an API, then that really does sound like the creators of the tool are at fault.

I think the important thing to note in both dog attack scenarios presented is that the owner is responsible in both cases. Either they purposefully created an unsafe situation or they were negligent in protecting the public from their property. Whether the dog is euthanized is about preventing it from happening again. Preventing it from happening in the first place is done by making the owner liable to disincentivize it.

Personally, i don't care about the end users. If you want to read my source i welcome that. I just want the CoPilot model and system open, since it was based (in part) on my work. Otherwise they are free to remove my work.

What was your plan when someone eventually infringed on your work?

If you want people to abide by your license, you have to enforce it yourself.

Of course, but you will not face manslaughter charges in that case.

So, following the same logic, if you train your copilot NOT to infringe on other people's copyright and it still does, it should be destroyed no matter your intentions. But at least you won't be charged with copyright violation yourself.

That said, I don't believe Microsoft's actions to be benign. I think this copyright whitewashing scheme is fully in line with their old MO, purposefully creating a legal quagmire surrounding all open source code.

- whether there are any damages - what the big deal is

In my mind, the entire thread identified a lot of those, but I think someone already said that it will likely be tested in court ( and I have zero idea, which way it will turn ).

For the record, I personally think Copilot is a cool tool ( frankly, it is not that different from automated stack exchange in terms of results ). If I worry about anything, it is that the overall standards will decline even further.

If you took a bunch of copyrighted and non-copyrighted books, cut them into pieces, shuffled them all together, then picked a passage at random from a hat; "not knowing" what you are going to get doesn't mean you aren't violating copyright.

That's essentially what copilot is doing: it's taking a bunch of code - some of it copyrighted without license - and using it as a dataset. The ML algorithm then tries to pattern match against that data to provide the user with something they want. That's just copyright violation lottery with extra steps.

All the problems and confusions mentioned above is due to this concrete inherent property of the Copilot.

If Copilot was made to help rearrange [1] existing code to satisfy new or changed needs, there would be no need in such a deep and explanative analysis of yours.

[1] https://www.folklore.org/StoryView.py?story=Negative_2000_Li...

Code is a liability. Less code is less liability. New code is a new liability.

Even tools to create new code is a new and unknown liability, it seems.

Which laws are considered in this case? I understand that fair use is a US concept. For example how does that apply to my projects, published and licensed by a European living in a European country? I would expect the majority of GitHub contributors to not be based in the US, so what laws should be considered?

> Except to the extent applicable law provides otherwise, this Agreement between you and GitHub and any access to or use of the Website or the Service are governed by the federal laws of the United States of America and the laws of the State of California, without regard to conflict of law provisions. You and GitHub agree to submit to the exclusive jurisdiction and venue of the courts located in the City and County of San Francisco, California.

[0] https://docs.github.com/en/site-policy/github-terms/github-t...

The mechanisms limiting this are mostly about privacy. Not whether you can agree to adjudicate copyright or TOU in California

Some seem to assume there's some general "If it's American it's invalid" law in Europe. This is not the case. With the exception of specific laws, such as GDPR regarding privacy, this is a perfectly valid clause.

Which is what everyone is talking about here.

For criminal, there is only jurisdiction in Sweden if the crime happened in Sweden. I would need you to link me a case where Sweden criminally convicted someone for copyright infringement who wasn't Swedish and wasn't in Sweden.

For example, I am not Swedish and do not travel there. Sweden has no power to enforce its laws against me. No matter what I do, I shouldn't be able to be convicted of criminal copyright infringent in Sweden.

Additionally copyright infringement can be a criminal matter in my country and the Swedish prosecutors have certainly not signed these agreements.

CoPilot doesn't seem to be a terrible implementation, instead it seems to be relying on it operating in a grey area. So they're going for broke, to try and get wide enough adoption that it becomes a fait accompli.

Not really.

A human being learns by doing, it takes a lot of time, their knowledge is not transferable, and, above all, they buy the material they learn from (most of the time) It's not fair use, it's "I paid for the entire opera" , sometimes multiple times: different editions, movies, tv shows, etc.

Secondly, it's not true that derivative material is automatically copyright free.

It is in all honesty the contrary, most derivative work that reached popularity is plagued with plagiarism, lack of attribution, undisclosed ghost authors etc. all things that get settled with a contract or in court if the publisher thinks it's worth it.

Otherwise the publication simply disappears.

In other cases the work is licensed, so that the publisher can use someone else's IP and literally resell other people's ideas and/or change them the way they like (or the license permits), without having to create new material and take the risk that nobody will notice it.

Case in point (among too many)

https://en.m.wikipedia.org/wiki/Legal_disputes_over_the_Harr...

This actually depends on how you train the model. Techniques such as using unlikelihood to penalize plagiarizing models exists. Microsoft/OpenAI are of course aware of those techniques but have chosen not to use them. The reason why is not difficult to figure out. Because the model hasn't learned how to implement sparse matrix multiplication in C, it has learned how to spit out someone else's code with a few variable names changed. Not unlike how many CS students not cut out for software development try to pass their entry-level programming courses. Professors use anti-cheating software to catch cheating students. Such software would catch Codex too and expose it as incompetent. Hence why it is not used.

Imo it is very simple: IP law is intended to incentivize creative work, so that it remains possible to profit from one's creation in an environment where it might be easier to copy than it is to create. We just need to figure out what outcome we want to create: one which incentivizes human creations, or AI "creations" - and build a legal framework to support it.

"Old systems suck and our new system is great and it is new technology. Therefore old rules do not apply to it."

Not surprisingly, the moment crypto started gaining traction, everyone was quickly made to understand that rules do indeed apply even if it is a new a facet of finance regulations ( or in the case of Copilot copyright law ).

For the record, I am sympathetic to your sentiment, but you can't really expect existing interests to accept a major change if it happens to undermine someone's way of life and, possibly, alter the current legal landscape. And this may end up a much bigger change than expected and may finally usher in the era management always wanted.

This doesn't seem to translate well to code. You can't copyright 'rembrandt's style', which is what dall-e and co learn from analysing those paintings. But what copilot gives you is a sizable chunk of code. That code is (mostly) exactingly precise: It's not like the AI learned a style and recreates the style. It learns what you intended to do and then verbatim copies a code chunk in. I'm pretty sure the AI part comes in to determine what it is you were likely trying to do, so that it knows which code to copy. Not to generate the code out of the AI model. At least, that's how I understand it works.

That is the fundamental difference.

> OK, if you said "what word comes next? MR. AND MRS. DURSLEY OF NUMBER FOUR PRIVET DRIVE WERE PROUD TO SAY" ... same thing

Let the AI fill in the rest of that sentence and we can debate whether that is copyright infringement or not. However, if the AI system is capable of finishing that sentence, it can presumably fairly trivially be asked a slightly different question. Instead of 'finish the sentence', how about: "Suggest the next likely sentence". That system would presumably generate the next exact sentence straight from the book, and keep going and - voila you recreated the entire book.

Which is clearly copyright infringement.

AI systems have a 'volume dial' to configure how much they mix and match. Turn it down low enough and asking DALL-E for 'a girl with a blue bandana and an earring in the style of Vermeer' will just give you a copy of Girl with the Pearl Earring, reproduced sufficiently accurately that it's trivially a copyright infringement (let's leave out the notion that the painting has moved past its copyright date of course).

Point is, for copilot, the volume dial has to be kept extremely low, because you can't just mash 5000 snippets together, unless those snippets are identical. Which is its own intriguing copyright infringement conundrum (500 artists each individually paint the same thing, and they can each prove they weren't influenced by the others, thus, not copyright infringement. Then, you reproduce the averaging of the 500, which results in yet another painting. Did you just infringe copyright? Surely the answer is 'yes', but whose copyright did you infringe? All of them? Is it 'yes'?)

You have to be very careful with this line of thinking. I remember SCO versus RedHat began on much smaller premises. I also remember it took years for ReactOS to audit their code after mere suspicions arose about code that seemed to be inspired by something like asm to C translation.

The GPL is clear: derived works must be under the GPL too. The license must be respected, it doesn't matter if it was copied "from inspiration because it was learned" by an algorithm or a person.

Many of humanity's best works (paintings, classical music, golden age of physics) have been created before humans voluntarily reduced themselves to automata.

AI hasn't produced anything apart from mashing together other people's creations, usually with a somewhat creepy result.

In programming this may work because quality does not matter, only LOC and social capital with the rest of the brogrammers. The objection that therefore real programmers do not have to be afraid is false. They either have to join the mediocrity or clean up the mess that the brogrammers make (while being disrespected by them of course).

The problem here is the scale and the power that enables that scale. This is industrial level mining of non-consenting humans, exploiting their life's work in many cases.

Yes, but a human being isn't allowed to copy that work before learning from it, even if they destroy the copy afterward. AIs don't watch Youtube or browse Github. People download copies of content stored there, analyze and categorize it, and then feed it into AIs. Copyright is broken at step 1.