| There's a couple of related issues being conflated here, and I'm not sure which one to bring up, mainly because I'm not sure in what direction the copying would be ruled to have gone. So I'll just mention all the cases. The first thing to note is that nonliteral copying can still be infringing. Actually, among copyright cases that actually go to trial, most of them are not bit-exact matches ("striking similarity" in legalese). The lower standard those cases would have to meet is substantial similarity, which requires proving both access and similarity. In other words, in order for you to produce a copy[0], you have to have both seen the original and produced something that is close enough if you squint at it. So let's say Papermark is the original and Corgi Dataroom copied it. If this actually went to trial, a significant amount of discovery would be spent harvesting all the e-mails and messages Corgi's development team sent to one another. Any evidence of knowledge or access to Papermark would probably be enough to prove a copyright violation. You mentioned LLMs, and some of the tweets here also mention them. I have no clue if either product used an LLM, but it's important to note that in the US, anything written by an LLM does not accrue copyright protection. So, if Papermark was LLM-authored, as a threshold matter, they would have to register[2] a very specific copyright that neatly carves out the LLM-authored bits. The judge would then only consider the parts of the code with verifiable human authorship, which would severely weaken Papermark's case. In the reverse case - i.e. Corgi Dataroom is LLM-authored - then Papermark's case becomes stronger. Note how I didn't say "AI slop is public domain" last paragraph, because it isn't. There are still unsettled legal questions as to whether or not training on copyrighted works is legal and if using an LLM trained on that work constitutes access to it. Furthermore, LLMs can have search tools that would give them access to code not within the training set, which would also be a more straightforward copyright violation. So if it turns out Dataroom's developers are all Claude fans, and Claude is copying Papermark code, then it's just a normal license violation. LLMs do not launder copyright. We can also consider the case where BOTH tools are LLM-authored. In this case, there might just not be a copyright case at all. Technically speaking, there would be some third class of plaintiffs who have been infringed, but they would have to choose to sue. It is a long-standing principle in law that you are not allowed to sue for other people's harms[1]. So in this case, nobody would have a case. > Now software developers are feeling what authors and artist felt It was specifically the FOSS community that sounded the alarm about training data theft first, because the FOSS community correctly understood coding agents to be an attack on copylefts[3] (albeit through the incorrect belief that LLMs could launder away copyright interest, as opposed to it just being told to make a noninfringing substitute). [0] I am skipping over notions of fair use and derivative works as they would complicate the analysis and do not apply here. [1] In general, the operating principle of American courts is "fuck around and find out", and this implies that the court is only allowed to find out once someone has fucked around. Otherwise, the courts could just sue themselves to rule on whatever the hell they want. Isn't adversarial common law GREAT!? [2] Yes, copyright registration is mandatory in the US, otherwise you can't sue, which is the whole point of copyright. The Berne Convention only half-applies here. [3] Clauses in licenses that require modifications to the work to be provided under the same license terms. Creative Commons calls these Share-Alike licenses. |
That's an easy case, but a more interesting question to me is if there wasn't any direct source code access (similar to an open source example I can't remember the name of atm). If you give an LLM requirements matching the copyrighted products capabilities, and it doesn't have Internet access, and it spits out different code that accomplishes the same stuff papermarks product does, do you believe courts would allow that, and copyrighted could be laundered legally?