New York Times vs OpenAI: grab your popcorn

[golden_labels]

OpenAI: ‘The New York Times Paid Someone to Hack Us’

The moment I feel no positive feelings towards either side and can just comfortably sit in the chair and watch the spectacle.

[jpanhalt]

I've knowingly had very little contact with AI, unless the annoying robo telephone trees one encounters in almost any business call today count.  As for copyright/copyleft infringement both fair use of copyright and wholesale reuse of copyleft require attribution.  It seems AI hasn't done that in the cases that make the news (e.g., legal briefs written by AI).  That's infringement. 

Since copyright claims have to be Federal in the US, that's the only recourse one has, and I don't blame the complainants.

[Marco]

It's a shame statutory infringement is only available to parties with registered copyrights, but even that will probably be enough to kill OpenAI outright with how much registered content they copied into their training set if it's not ruled fair use.

Though I hope there isn't any judge who will countenance the argument that since any individual third party content on the internet is almost worthless, they should just be available for pirating by the AI companies. If so, once fair use is out of the window we can just all join the class action lawsuit for nothing rounded down.

[golden_labels]

Not any action performed in relation to a work is considered “using”, or requires attribution or even permission. In particular if the work is publicly displayed, as it is in the case of data fed to recent LLM and image generation smortnets. To make a counterexample, nobody can require you to obtain permission to look at a publicly visible building (even in France, where there is no FoP!). Neither you have to make a proper attribution, when you describe what you have seen. Even if you do that commercially (a movie review, a travel guide, a map of an area etc.) I believe everybody agrees, these don’t constitute “use” of a work, despite some action and a kind of consumption of a work occurs. Of course courts and lawmakers may decide otherwise, but I can’t see it as a trivial case.

Considering a model itself to be a derivative work is playing with fire, though. It’s not a transformative action, that makes a work derivative, but presence of similar elements. A model does not contain them. So to consider a model being a derivative work, we would need to either make the transformative action or transformed (but not similar) content presence to be the condition. And this is opening a can of worms, because it widens the concept enormously. It’s easy to suggest it with relation to smortnets, but even easier to forget it will also apply to everything else.

Neither of the above relate to the output of the smortnets. The output may be a derivative work and infringe intellectual property. Like with any tool. Smortnets just seem to make this risk much higher to the user than other tools.

[magic]

Yep, there is nothing entirely obvious about this case except that lawyers will be the real winners there

I do find it funny how they claim that NYT "hacked" them, simply by manipulating queries to get the output they wanted.

[Marco]

Not any action performed in relation to a work is considered “using”

Every copy is considered a copy, it took a lot of lawsuits against mostly Google and the DMCA to carve out copyright exemptions on the internet ... and none of that directly applies here. So the copy into the training set ... yep, that's a copy. The way it's used only becomes relevant for fair use, that's really their only shot.

In particular if the work is publicly displayed

Public access is not public domain.

Neither you have to make a proper attribution, when you describe what you have seen.

Computers don't see anything, they copy and process. They're not humans.

Considering a model itself to be a derivative work is playing with fire, though.

It's besides the point really, if it's not fair use the statutory infringement for the copy into the training set will destroy them outright. There will be no money left for thinking about damages from derivative works.

PS. in theory EU law is different, in practice we're puppy dogs. So if the US supreme court deems it fair use, the EU laws will change to parallel.

[jpanhalt]

Not any action performed in relation to a work is considered “using”, or requires attribution or even permission. In particular if the work is publicly displayed, as it is in the case of data fed to recent LLM and image generation smortnets.

Copyrights and patents are determined by the country.  In the US, a copyright owner retains those rights even after a work is put for public display. Thus, copyrights exist in the US even for articles published in newspapers or movies shown around the world.  That's the main purpose in having them. In the US, copyright to a work is created at its creation regardless of whether the copyright is registered or the work is published.   China is quite different. 

[Marco]

Registration is important for statutory fines, no need to prove damages. Basically if it's not fair use, OpenAI is bankrupt regardless of calculation of damages. Microsoft might not do much better, though at least most source code isn't registered.

[SiliconWizard]

Yeah the question of current AI models being fed off copyrighted material without any consent has been going on for a while now.
Good luck claiming that "fair use" can be considered for multi-billion-dollars businesses based on unauthorized copyrighted stuff for a very large part. Almost absurd.