Hacker News Reader

> The copilot team rushed to slap a copyright filter on top to keep these verbatim examples from showing up, and now claims they never happen.

More than that: the fact that they claimed it wasn't possible before adding the filter, to filter out the thing that said wasn't possible. This doesn't help me trust anything else they might say or have already said.

My take on that was always: if it isn't possible, then why are MS not training the AIs on their internal code (like that for Office, in the case of MS with their copilot product) as well as public code? There must be good examples for it to learn from in there, unless of course they thing public code is massively better than their internal works.

How do you know they aren’t training it on their internal code?

Since you really need to work hard to make the AI spit out anything verbatim, and you have no knowledge of their internal code, how could you ever prove or deny it?

> How do you know they aren’t training it on their internal code?

Because if they were, they would have said.

It would be an excellent answer to the concerns being discussed here: “we are so sure that there is nothing to worry about in this regard, that we are using our own code as well as the stuff we've schlepped from github and other public sources”.

> Just because you filter out verbatim copies doesn't mean there isn't still copyright infringement/plagiarism/whatever you want to call it.

Actually, it does. The production of the output is what matters here.

If you copy someone else's copyrighted work and then rearrange a few lines and rename a few things, you're probably still infringing.

For a book or a song, for sure, although that isn't really punished. Search the drama surrounding a popular YA author in the 10's, Cassandra Claire. For code since you can only copy the form and not the function that might actually be enough.

People do clean room implementations because of paranoia, not because it's actually a necessary requirement.

Moving a few things around means your internal process already had copywrite infringement.

Probably not. Copyright infringement in the manner we're talking about presumes you already have license to access the code (like how Github does). What you don't have license to do is distribute the code -- entirely or not without meeting certain conditions. You're perfectly free to do whatever naughty things you want with the code, sans run it, in private.

The literal act of making modifications isn't infringement until you distribute those modifications -- and we're talking about a situation where you've changed the code enough that it isn't considered a derivative work anymore (apparently) so that's kosher.

First the case would be dismissed if Copilot had permission to make copies. Clearly they didn’t. Copyright cares about copies, for profit distribution just makes this worse.

> you already have license to access the code

This isn’t access, that occurs before the AI is trained. It’s access > make copy for training > AI does lossy compression > request unzips that compression making a new copy > process fuzzes the copy so it’s not so obvious > derivative work sent to users.

Clearly Copilot had permission to make (unmodified) copies, the same way Github's webserver had permission to make (unmodified) copies. The lawsuit is about making partial copies without attribution.

GitHub's terms of service (TOS), in my non-lawyerly opinion, clearly states the license for uploaded works granted to them by users doesn't cover using the data to train an LLM or any kind of model beyond those used to improve the hosting service:

>You grant us and our legal successors the right to store, archive, parse, and display Your Content, and make incidental copies, as necessary to provide the Service, including improving the Service over time

>This license does not grant GitHub the right to sell Your Content. It also does not grant GitHub the right to otherwise distribute or use Your Content outside of our provision of the Service, except that as part of the right to archive Your Content, GitHub may permit our partners to store and archive Your Content in public repositories in connection with the GitHub Arctic Code Vault and GitHub Archive Program.

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

I think the important questions are (1) whether "the Service" includes Copilot, and (2) whether GitHub is selling users' content with Copilot.

For (1), I'm unhappy to admit Copilot probably does fall under "the Service," which is nebulously defined as "applications, software, products, and services provided by GitHub." But I'll still say that users' could not agree to this use while GitHub was training The Copilot model but hadn't yet announced it. At that time, a reasonable user would've believed GitHub's services only covered repository hosting, user accounts, and the extra features attached to those (issue trackers, organizations, etc).

GitHub could defend themselves on point (2) by saying they aren't selling the code, instead selling a product that used the code as input. But does that differ much from selling an online service that relies on running user code? The code is input for their servers, and it doesn't need to be distributed as part of that questionable service. But it's a clear break from the TOS.

GitHub’s web server is not the same thing as Copilot and needs separate permission.

GitHub didn’t just copy open source code they copped everything without respect to license. As such attribution which may have allowed some copying isn’t generally relevant.

Really a public repo on GitHub doesn’t even mean the person uploading it owns the code, if they needed to verify ownership before training they couldn’t have started. Thus by necessity they must take the stance that copyright is irrelevant.

If you’ve copied three lines and rearrange and reword them, there’s little infringement left.

If you copy a whole book and do the same, there’s still lines-3 infringement left.

> 1.

Isn't that akin to destruction of evidence?

Legally? No.

In spirit? ... Probably?

Unlike most LLMs, Github copilot can trivially solve their copyright problem by just using only code they have the right to reproduce.

They have a giant corpus of code tagged with license, SELECT BY license MIT/Equivalent and you're done, problem solved because those licenses explicitly grant permission for this kind of reuse.

(It's still not very cash money to take open source work for commercial gain without paying the original authors, and there's a humorous question if MIT-copilot would need to come with a multi-gigabyte attribution file, but everyone widely agrees it's legal and permitted.)

The only reason you'd hack a filter on top rather than doing the above is if you'd want to hide the copyright problem. It's an objectively worse solution.

> Unlike most LLMs, Github copilot can trivially solve their copyright problem by just using only code they have the right to reproduce.

Absolutely not trivial, in fact completely impossible by computer alone. You can't determine if you have the right to reproduce a piece of code just by looking at the code and tags themselves. *Taps the color-of-your-bits sign.*

* I can fork a GPL project on Github and replace the license file with MIT. Okay to reproduce?

* If I license my project as MIT but it includes code I copied inappropriately and don't have the right to reproduce myself, can Github? (No) This one is why indemnity clauses exist on contracted works.

* I create a git repo for work and select the MIT license but I don't actually own the copyright on that code and so that license is worthless.

There is no difference when it comes to MIT and GPL here. If your model outputs my MIT licensed code, you still need to provide attribution in the form of a copyright notice as required by the MIT license.

Have the copyleft people, or anyone else, produced some boilerplate licenses that explicitly deny use in training models?

I would think it is pretty obviously not.

Is taking away a drunk driver's keys (before they get in the car) destruction of the evidence of their drunk driving?

This is not what I meant. By placing a copyright filter and claiming it never happened (please read the line I was replying to) before the system can be audited, they're indeed taking away the drunk driver's keys, which is a good thing, but also removing the offending car before Police arrives.

In this metaphor, removing the car of someone who was going to drink and drive but didn't, is certainly not a crime. Presumably though you mean removing the car after drunk driving actually took place - which might be, but probably depends a lot on if the person knew, and what the intent of the action was.

In the current case - its unclear if any crime took place at all, it seems clear that the primary intent was to prevent future crime not hide evidence of past ones. Most importantly the past version of the app is not destroyed (presumably). Github still has the version of the software without the copyright filter. If relavent and appropriate, the court could order them to produce the original version. It can't be destroying evidence if the evidence was not destroyed.

Yes, sorta. We're talking about software, therefore a piece of code that does something programmatically isn't like the drunk driver in a car that may cause more accidents, and although we aren't sure about that we prevent him/her to drive anyway just to be safe. The software would most certainly repeat its routine because it has be written to do so, that's why I wondered about destruction of evidence; by removing/modifying it, or placing filters, they would prevent it from repeating the wrongdoing, but also take away any means of auditing the software to find what happened and why.

Not in any way I'm aware of - and would be required if they were served a DMCA notification/Cease and Desist against a specific prompt.

The people that think Copilot is infringng their copyright would be happy with that I would think? Unless they take a much stricter definition of fair use than current courts do.

No more so than scanner/printer manufacturers adding tech to prevent you from scanning and printing currency is destruction of evidence that they are in fact producing illegal machines for counterfeiting.

> The copilot team rushed to slap a copyright filter on top to keep these verbatim examples from showing up, and now claims they never happen.

Well if the copyright filter is working they indeed aren't happening. Putting in safe gaurds to prevent something from happening doesn't mean you're guilty of it. Putting a railing on a balcony doesn't imply the balcony with railing is unsafe.

> LLMs are prone to paraphrasing. Just because you filter out verbatim copies doesn't mean there isn't still copyright infringement/plagiarism/whatever you want to call it

Copyright infringement and plagerism are different things. Stuff can be copyright infringement without being plagerized, and can be plagerized without being copyright infringement. The two concepts are similar but should not be conflated, especially in a legal context.

Courts decide based on laws, not on gut feeling about what is "fair".

> They clearly know the problem is real

They know the risk is real. That is not the same thing as saying that they actually comitted copyright infringement.

A risk of something happening is not the same as actually doing the thing.

> "Ner ner ner ner ner, you can't prove it to a boomer judge".

Its always a cop-out to assume that they lost the argument because the judge didn't understand. I suspect the judge understood just fine but the law and the evidence simply wasn't on their side.

> Well if the copyright filter is working they indeed aren't happening. Putting in safe gaurds to prevent something from happening doesn't mean you're guilty of it. Putting a railing on a balcony doesn't imply the balcony with railing is unsafe.

Doesn't mean you weren't, at some point, guilty of it, either. It doesn't retcon things.

Sure, which is why we require evidence of wrong doing. Otherwise its just a witch hunt.

After all, you yourself probably cannot prove that you didn't commit the same offense at some point in time in the past. Like Russel's teapot, its almost always impossible to disprove something like that.

Yeah but I think the main concern in this situation is copilot moving forward, not their past mistakes.