Hacker News Reader

Hacker News

Comment by pmlnr | Hacker News Reader

Parent Full thread

pmlnr

It's not "fair-use classification of ML models". If the ML black box:

- produces an output

- that output is the exact same thing - text, code, image, et c - that went in

- that thing is licensed

- fails to fulfill the requirements of the license

than it's breaking the license. I don't see how it can be wiggled around.

The black boxes that produce entirely different outputs, or the ones that slurped billions of photos without even blinking at license but never vomits actual images are, at least, not reproducing the source material.

View on HN

I think another important analogy is that you couldn't do this same operation with a human. In my eyes a human and an ML algorithm are both "agents". An ML algorithm is a computerized actor on a system but they are still able to do something novel based on previous inputs. If we take these two scenarios:

1. A hypothetical employer asks an employee to find a really fast algorithm to find an inverse square root. The employee has a perfect memory and can regurgitate fast inverse square root from Quake 2.

2. A hypothetical employer asks an employee to find a really fast algorithm to find an inverse square root. The employee uses this ML algorithm to generate the code which essentially regurgitates the only previously seen inverse square root function it was trained on.

At some point in time in the future id Software sees this is happening at our hypothetical company and sues. In neither situation is the "actor" held accountable since the "actor" was employed by the company which allowed for practices that enabled this infringement to take place.

I think an important legal question is: Do we consider a ML algorithm being trained on data the same as a human reading prior art for an innovation?

> I think an important legal question is: Do we consider a ML algorithm being trained on data the same as a human reading prior art for an innovation?

It seems pretty obvious that it is the same thing to me. Instead of the human reviewing the prior art themselves the human has written a program to do so, but this can't change the acceptable outputs produced by the human's creative process (that includes the ML in this case).

I agree with this sentiment but I don't know if the courts do. There is so much cruft in the legal system from problems entirely removed from today's moral and ethical standards for disagreements from bygone eras that will be used as case law here.

If people can write programs that launder IP then IP is over. That is never going to fly in any court.

It's more complex. Are you laundering 5 lines of code? Probably fair use. Are you laundering Windows? Clearly not okay. Are you trying to launder, or is it unintentional? Etc.

There's a line somewhere, but it's somewhere, not at either edge, and it's a complex, odd-shaped line we haven't yet defined. I don't know where it is. However, it's not where you're putting it, and there isn't a slippery slope to "IP is over." If you use five lines of my code in a different system, IP hasn't ended.

My point is only that running the text through an ML algorithm changes nothing. If the 5 lines are fair use, you could have copy pasted the lines yourself. If its infringement, it doesn't matter which tools you used to put the infringing code in your file.

> Are you laundering 5 lines of code? Probably fair use.

The amount of copying does not matter for a copyright claim. If you copy a single character from a codebase that could could get nowhere but there, and lawyers could prove it, that could go to court. This is a hypothetical but this is entirely possible. There have been copyright cases fought over single sentences especially in the music industry.

> Are you trying to launder, or is it unintentional?

I'm not a lawyer but in the US, Copyright is a strict liability statute which means intent does not matter.

>> Are you laundering 5 lines of code? Probably fair use.

>The amount of copying does not matter for a copyright claim.

This is incorrect. The amount of material used is the third factor in a fair use test. There are other ways it comes up as well (e.g. damage calculations).

>> Are you trying to launder, or is it unintentional?

> I'm not a lawyer but in the US, Copyright is a strict liability statute which means intent does not matter.

In this context, this is incorrect as well. People seem to be misinterpreting laws pretty badly, so instead of explaining how to apply here, I'll give a simpler, analogous context:

https://en.wikipedia.org/wiki/Online_Copyright_Infringement_...

In this context, it depends on a lot of things, such as direct versus contributory infringement. This would likely be a contributory infringement case, where intent is almost a requirement. https://www.legalmatch.com/law-library/article/what-is-contr...

Keep in mind these are all factors. They matter, so they'll help swing a case one way or the other, but they're not something you can bank on in isolation. That's what will make for interesting case law.

> This is incorrect. The amount of material used is the third factor in a fair use test. There are other ways it comes up as well (e.g. damage calculations).

The 3rd factor of the fair use test deals with how substantial the copied material is in relation to the entire work. You can make a copy of 100% of the original work and have it still be fair use while in another situation copying 0.002% of an original work would not be fair use. The measurable amount (bytes, seconds, square inches) you are copying does not matter here. Instead the "importance" of what you are copying to your criticism is being defined in this prong.

For instance:

I am a movie buff and:

1. Make a copy of the entire movie of Citizen Kane 2. Remove the audio and replace it with a commentary track going over every framing/camera trick through the movie 3. Upload this to an educational youtube channel that teaches viewers how to shoot a movie

I would likely be allowed to use fair use as a defense to a copyright claim. Here, if an expert could say all 100% of Citizen Kane's film, contained relevant film techniques that I was actively commenting on, I would have copied the correct amount for my usage.

If I instead owned a movie review channel and I said "Citizen Kane is my favorite movie" and I displayed the entire movie after that I would likely not be using an appropriate amount of the film.

It's also important to note that fair use is a very narrow defense that covers a very limited subset of uses. It is not a generally applicable, or even a guideline, for how to skirt around copyrights.

There are plenty of errors in this, and you're confusing the four prongs, but I'll run with it for now, since it's good enough.

If I am doing code autocomplete, and copying three lines out of your program, those three lines are exceptionally unlikely to be "substantial ... in relation to the entire work."

Where you're confused is lots of places, but the biggest one is you're mixing up the prongs. Most of the reason your Citizen Kane example might work (it probably wouldn't) is Factor #1, character of use. Commentary and education are favorable.

I can come up with examples where 0.002% of an original work would not be okay, but they're pretty contrived.

For your benefit, a random link: https://guides.lib.utexas.edu/fairuse/fourfactor

Size definitely matters. It's not a hard-and-fast rule, but a factor.

The whole reply I post here is made from letters that I copied out of a Harry Potter PDF.

Clearly that's not how copyright works.

Programmers generate every possible melody in MIDI to prevent lawsuits - https://news.ycombinator.com/item?id=22440944

I think you missed the part where he states:

> that could could[sic] get nowhere but there, and lawyers could prove it

Also, the letters in this case are not representative of the creative work itself. If you used the title font to write your post (hypothetically, I know HN does not allow this), you'd be in bad water already.

I could copy a whole sentence from the book and it would not matter at all. Or even 5 lines like the parent post said.

>....you'd be in bad water already.

Not true either, writing my message in harry potter font would do absolutely nothing. the internet is full of images which use these or similar looking fonts. Its called fan art or whatever, copyright laws are not applicable to this.

Exactly zero real-world situations will have no "substantial similarity" (that's the legal requirement - https://en.wikipedia.org/wiki/Substantial_similarity ) except for a single character.

Please explain how it "is entirely possible".

Bear in mind that there must be creativity in that choice of character to qualify for copyright protection. If only one character is correct, there's no creativity.

A sentence in a song has far more discretion for creativity than a single character in a software program.

As best as I can tell, you would need to construct some sort of super-APL where "a single character" had significantly more information content than any Unicode glyph, to exceed any de minimis standard applied to a programming language.

(Think Prince's "Love Symbol".)

> that could go to court

I think you mean "that could be found infringing." Going to court is trivial, even if there is no infringement.

> As best as I can tell, you would need to construct some sort of super-APL where "a single character" had significantly more information content than any Unicode glyph, to exceed any de minimis standard applied to a programming language.

This used to be something cartographers did: https://en.wikipedia.org/wiki/Trap_street

Even if you didn't have the street's name, if you copied the road on the map (no words, just the shape) they would know you stole their IP. I do not think there is literally a character that would satisfy this condition. The point of what I was saying was the "size" of the copying does not matter at all.

The Substantial Similarity is explicitly about this phenomenon. For something to be substantially similar, from a software perspective, you could hypothetically see the modules and data structures contain roughly the same data types and that the flow of logic is the same. You could also point to a function and say "these are the same variable names". Substantial similarity is not the only component of a legal analysis of a copyright claim. Another important component of copyright is if someone had access to the material. If I had code on my laptop that I showed to no one, uploaded nowhere, and at some point I find another developer who is doing some code that is character for character identical to my work, I likely cannot do anything.

The law is complex here, I am not a lawyer, but I think the important component here: The length of copying is not a factor at all. If for some reason your algorithm requires `magic_code.seed(20154)` and your code does this by doing `magic_code.seed(ord('人'))` and you find someone else doing the same thing you'd definitely want to investigate what's going on there.

The entire copyright doctrine is created to help protect the creativity of an author. If there is something creative in your code that is copied, you likely have a copyright claim you could argue.

Your scenario was "If you copy a single character from a codebase that could could get nowhere but there". It was that "entirely possible" scenario I questioned.

Not "roughly the same data types", "same variable names", "magic_code.seed(20154)", etc. I have no question with those. Those are enough characters that they may contain creative - and thus copyrightable - content.

That is, I question your assertion that 'the "size" of the copying does not matter at all' by asking you to come up with an entirely possible scenario for why there is no de minimis case in software, even down to a single character, when there is in every other area of copyright.

BTW, the WP page you pointed to notes that "Trap streets are not copyrightable under the federal law of the United States." Thus, "stole their IP" has no meaning - they have no copyright, trademark, patent, etc.

> I question your assertion that 'the "size" of the copying does not matter at all' by asking you to come up with an entirely possible scenario for why there is no de minimis case in software, even down to a single character, when there is in every other area of copyright.

Where in copyright law for any area is there an established minimum "length"? I've never heard this nor heard lawyers claim of such a thing.

I never made that claim. I even posited a "super-APL" where a single character was enough to contain the creativity which is a prerequisite for copyright protection. Such a language does not exist, and I don't think it ever could, but there is a slightly-greater-than-0% chance that it could happen.

OTOH, your original context for an "entirely possible" scenario at https://news.ycombinator.com/item?id=27730596 was "The amount of copying does not matter for a copyright claim", which implied your scenario was not that hypothetical example I posited.

(A 1E-20 probability * maximum expected lawsuit payment with successful lawsuit = don't worry about that scenario.)

We do know that the courts have decided many cases are de minimis non-infringing use of materials otherwise under copyright.

Do you think there is no acceptable de minimis argument in software? If not, why is software somehow special compared to other areas of copyright?

A "de minimis analysis ... usually focuses on the amount of the copyrighted material that is copied." - quoting http://patentarcade.com/2020/04/nba-2k-avoids-tattoo-copyrig...

If 'de minimis' use exists in software, what realistic scenario are you thinking of where a single character is enough?

To be clear, I think zero shared characters can still show copyright infringement, for reasons you mentioned about abstraction-filtration-comparison.

But your "entirely possible" example had no other shared similarities beyond a single character, and I can't see how any court wouldn't think that was a de minimis use of that character - assuming it had copyright protections in the first place!

> The amount of copying does not matter for a copyright claim.

It absolutely does for fair use analysis, see 17 USC § 107: “[…] In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include—[…]the amount and substantiality of the portion used in relation to the copyrighted work as a whole”

> I don't know where it is. However, it's not where you're putting it

How are these statements not contradictory? Do you know where the line is, or not?

I don't know where my airpods are. However, they're not on the table in front of me.

I know that my pizza is not dark matter

One big difference is that the human will (... hopefully) never be considered themselves a derivative work, whereas there seems to be a case to be made that the model is.

Algorithms are not subject to copyright. Writing and copying an inverse square root is fine, doesn't matter if it's human, computer or alien.

The trouble is verbatim copy of original copyrightable code, with an attached license that doesn't allow it. Github is most likely in violation because it's doing just that.

However that specific square root square example may or may not be a violation. There's a difficult question of what constitutes original copyrightable work, a trivial well defined algorithm isn't, a larger code with comments could be.

> Writing and copying an inverse square root is fine, doesn't matter if it's human, computer or alien.

This is very myopic because there's a big difference between "memorizing an algorithm" and "memorizing an implementation of an algorithm".

> There's a difficult question of what constitutes original copyrightable work, a trivial well defined algorithm isn't

I don't think copyright law agrees with this sentiment in the US. Even the spacing, variable names, and comments are non-functional attributes of this code. And, this also goes further, because the machine code it produces is also data that has copy rights. There's layers and layers of complexity here that is difficult to enumerate with forum posts.

>>> This is very myopic because there's a big difference between "memorizing an algorithm" and "memorizing an implementation of an algorithm".

No there's isn't. An implementation of an algorithm is not copyrightable. It's specifically cited as an exception in European directives.

I should say I speak more from a EU perspective. US may vary slightly. Both vary by country/state as well.

There is a challenge with determining the threshold for copyrightable original work and it's not the same in every jurisdiction. That we can agree on. (I personally think comments are a big problem because they're always freeform text). In my opinion the tool should consider doing an autoformatting and removing comments, to avoid these issues.

Here's a thread of insights on this from an actual lawyer: https://twitter.com/luis_in_brief/status/1410242882523459585

The gist:

>> "There is an observable trend in US law, based on fair use and older notions in US copyright law of the need for creativity, that judges give a looooot of leeway to “machines that read”. Copilot fits pretty squarely in that tradition."

It’s more of I created a search engine that is context aware. Then claiming that it generates novel code because it can take snippets from multiple sources. Then have a bunch of people not understand what it is then are trying to claim it does x which would make this a general ai.

It can be wiggled around. For example, if I ship a robot which can paint a famous painting, but not the painting itself, I would argue the copyright violation happens by the user of the robot when it paints the painting, not by the maker of the robot.

If that's something I'd planned on and intended, and you own the copyright to the painting, you have all sorts of legal tools to go after me -- contributory infringement, collusion, and so on. Still, contributing to a legal violation is not the same as engaging in one.

I suspect an argument can be made that github now incorporates AGPL code generated by co-pilot, and so is AGPL. A fair use argument might be made as well; we've all copied one-liners from blog posts and tutorials, and that's okay. I'm having a harder time seeing a reasonable argument that an ML model is directly infringing just because it can produce a copyrighted work, though.

So I don't think it's a cut-and-dry legal question.

I will give a caveat: Coders tend to read laws much too literally, like computer code. When I was was an obnoxious teenager, I thought I'd found all sorts of contract / license / etc. loopholes in all sorts of legal documents, and I considered myself profoundly clever, thinking I'd outsmarted the lawyers.

Nope.

In college I took law classes, got into the real world, did a few startups, and saw a few legal cases. Court systems have technical rules you need to be aware of, of course (e.g. if you miss a deadline....), but interpreting the law is really grounded in common sense. Common sense is culturally situated, and ours is situated in hundreds of years of case law.

>>> For example, if I ship a robot which can paint a famous painting, but not the painting itself,

Plot twist: Depending on the jurisdiction, you are responsible for distributing means specifically designed to circumvent copyright, which makes you liable for a much greater punishment, and liable for every reproduction done by your users.

There's a reason that painting robots are sold as general painting robots for various industrial usage, they're not advertised to and shipped with templates to reproduce copyrighted paintings. ;)

> if I ship a robot which can paint a famous painting, but not the painting itself, I would argue the copyright violation happens by the user of the robot when it paints the painting, not by the maker of the robot.

How is that different than sending someone a jpeg of a copyrighted painting? Or the jpeg and the executable that can render the jpeg. Or the jpeg and the executable and the computer it runs on. Each of those gets closer to what you said.

If the user of the robot is responsible for the copyright violation, then the user of Copilot is responsible, correct?

If that's the case, then there will not be any users of Copilot except those that unknowingly open themselves to liability. I suspect there will be few.

I would be okay with that situation. But I'm not okay with the situation as it currently stands.

I'm okay with where it stands.

Where it stands: We have a new technology. We don't really understand it. There aren't well-defined ethical or legal bounds. We're figuring them out, blumbering along as we go. People are trying things, and are making all sorts of arguments, some of which will stand the test of time, and some of which won't.

I'm more concerned when we do this with weapons, medicine, or social science, but an autocomplete tool seems like a relatively benign place to start figuring this out.

I don't want to agree with you that an autocomplete tool is a relatively benign place to start, mostly because I am upset about GitHub using my code in this way. But you are correct.

But let's fix this before it goes further, especially since mindlessly copying output from a machine will probably result in more bugs.

> copyright violation happens by the user of the robot when it paints the painting, not by the maker of the robot.

Thats very interesting. Because AFAIK Github is not "shipping" copilot to anyone. They run it and make it available only to those that use their online editor.

Unlike humans, ML models can be attacked to pull out the original training data.

Not all networks are susceptible to this, but if you can get a 1:1 copy, that's no different from bittorrent, limewire, or KaZaA. That doesn't seem implausible for text.

It's very different. Law looks at intent.

Napster was an easy litigation mostly because of intent. There was a document trail a mile long that the primary goal of Napster was to facilitate copyright infringement, although on paper, it could be used to share anything.

Cameras, xerox machines, and audio recorders can all be used for infringement, but aren't illegal since that's not their primary purpose. For even finer lines, see the Betamax case. You can also look at cases about different sorts of transformations and transmissions:

- I buy a DVD. It is transferred via USB cable from a DVD player connected to my computer, temporarily copied to computer memory, and then shown on my monitor.

- I buy a DVD. It is transferred via the internet from a DVD player in my data center, via the Internet, temporarily copied to your computer memory, and then shown on your monitor.

Can you see how these have nearly identical technology, yet different legally? If you ever take a law class, there are a lot of discussions about how to draw such lines (disclaimer: IANAL, but I've sat in on law classes before).

To be clear, I'm not taking a stance about where this ought to land legally. I don't know. But the analogy to bittorrent, limewire, or KaZaA is a faulty one.

GitHub and Microsoft intend to commercialize Copilot. This means they are taking huge swaths of copyrighted material that they did not license and using it to sell a commercial product. That really hurts their “fair use” defense. It’s even worse since they are facilitating the customers to pirate copyrighted material as well, which doubly harms their case.

I mean honestly I'm a black box and if I spit out the same code license and all, people have a problem with that.

> It's not "fair-use classification of ML models". If the ML black box: > [four part test for rejecting fair use that has two obvious conflicts with the nature of the fair use exception deleted]

What case did this proposed legal test come from?

Sounds like it's high time to abolish intellectual property, or at least software licensing and patents.

On the other hand though it could be argued that a 5-line snippet used in a novel context and/or modified is fair use.

Reply on news.ycombinator.com