Comment by user5994461 | Hacker News Reader

Comment by user5994461 | Hacker News Reader

Parent Full thread

99% sure that Nat Friedman is wrong there.

He's correct that AI can train, but it's not relevant.

(1) The issue is with CoPilot outputting copyrighted code AND attaching a separate license to it (looks like it's hiding the initial license and can generate another license automatically). This is reproducing copyrighted code and misrepresenting the license, neither is allowed.

(2) He posted that the output belong to the operator. It's factually wrong. Original copyrighted code belongs to their original writer, not the operator. In effect Github Copilot cannot "launder" license and it is very wrong for their CEO to claim otherwise.

(3) Looks like he's trying to waive responsibility of Github by stating that the developer receiving the code is responsible? Wrong, GitHub is responsible for their actions (vary with the jurisdiction). There's a complex matter of who's responsible when some proprietary code will end up in a company product and they get sued. The company is in violation, they can turn against GitHub for providing copyright code, GitHub was responsible for it. There's a complex chain of responsibility, any lawyer worth their salt would cringe at the claims and responsibility that GitHub is exposing itself to.

The thing that gets my goat here is that repos, and individual files, have their licenses. With effort, github could have linked snippets to the files they came from, and used a mostly-automatic process to identify the license and copyright holder of every quoted snippet. Then, as code gets generated, a conglomerate license could be assembled in parallel. It could emit warnings that licenses are incompatible, or better yet, give the user a choice over which licenses they find acceptable and disallow incompatible licenses.

This whole thing could be done above board, and it's pure sloppiness on the part of github that they've chosen not to.

> sloppiness

That’s an understatement. There is a deep lack of understanding of IP issues as demonstrated by their public representatives. Surprising for a company like GitHub where software IP should be a part of their core competency. They have made defenses like “it’s public data” which means nearly nothing in regard to copyright enforcement. The Avengers is public “data” (it’s not data in terms of the law), that doesn’t mean you can use the content for whatever you want. It shows a total lack of respect for the work that the FOSS community does and their rights. This has irreparably changed my perception of GitHub forever in a strongly negative way.

As I've said elsewhere:

> It shouldn't do that, and we are taking steps to avoid reciting training data in the output.

Seems like he agrees that __reproducing__ copyrighted code is not permissible.

Also seems likely to me that Nat etc were probably misled / incurious (no excuse) about precisely what CoPilot would do in practice.

> With effort, github could have linked snippets to the files they came from, and used a mostly-automatic process to identify the license and copyright holder of every quoted snippet.

Actually, they should not be able to, as the whole point is that CoPilot does not just paste code, but instead understands it and applies knowledge, not finished snippets.

The fact that it does paste snippets is the problem, not the authorship of the snippets, as these simply should not exist.

On one hand, there's a whole field of study on explicable AI and I'm unconvinced that this cannot be baked into that automatic reasoning process.

On the other hand, a wholly independent system to search the training set based on AST matching and text similarity could solve this problem in the majority of cases.

> pure sloppiness on the part of github

Considering that their entire business model from the start has been "capturing the open-source ecosystem onto a closed and proprietary platform that doesn't interoperate", I'm more inclined to assume malice here than 'sloppiness'.

Edit: By which I mean that their approach most likely just couldn't work well enough for end users, in terms of hassle, if they actually did it above board and included licensing information. So not caring about that was the profitable option.

> It shouldn't do that, and we are taking steps to avoid reciting training data in the output.

Seems like he agrees with you that __reproducing__ copyrighted code is not permissible. Doesn't answer the broader point about training with that code.

Whether or not the AI can be trained on GPL'd code without becoming GPL'd itself is completely relevant. That's what the article is about.

If it can then there's no problem unless CoPilot actually did produce code that was a close copy of GPL'd code and GitHub used it. Their own analysis says that that is really unlikely and easy to prevent anyway.

Reply on news.ycombinator.com