Comment by handoflixue
13 days ago
Regardless of what you do now, I think you should be prepared for the upcoming reality that LLMs are going to be able to reproduce software, feature-perfect, in a way that does not currently violate copyright law.
Right now, the settled law is that such an LLM reproduction is 100% legal.
If you really want to protect your software in the years to come, you might have to seriously consider starting some sort of popular political movement to address this issue in copyright law.
Current models can already do a full reproduction of anything with source code available (e.g. JavaScript...), and there's already been some poor-quality Photoshop knockoffs.
I feel like I need to nitpick a little: Models don't need source code available to reproduce software. See all the "full decompilation" projects cropping up. There's no putting this genie back in the box, because LLMs can also "refurbish" a project enough that it ceases to look like the original. People don't bother now because they don't have to, but in a world where they'd get hit with copyright notices, they would.
I will say that a lot of the "full decompilation" projects popping up are honestly crap - some are good, but a lot of them just build something that makes for a good screenshot or promotional video.
I'm not convinced we are actually at the point where something like Photoshop is trivial to rebuild. That involves a lot of manual QA and the expertise in actually knowing how everything should work.
That said, give it 6-12 months and I won't be surprised if they can one-shot "create a future-complete clone of Photoshop, make no mistakes"
> Right now, the settled law is that such an LLM reproduction is 100% legal.
How so? Interfaces are not copyrightable, but that it not the same as dissecting a js bundle and copying the implementations. Are we sure these LLM are keeping sanitary habits there?
if you've used any Ai in your own code authoring, copyrights may be completely out the window
several courts have ruled Ai output is not copyrightable, I am unaware of any co-authored cases
If you only used the output of an LLM, then you don’t qualify.
But, the use of LLMs is not disqualifying. To qualify for copyright protection your work simply must have a sufficient degree of human authorship.
However this is just about protection, not infringement.
If you use an LLM to generate something and that LLM just happens to output something that another human wrote, you may be liable for copyright infringement.
2 replies →
That doesn't top people producing copies though, just trying to copyright the copies.
1 reply →
> popular political movement to address this issue in copyright law.
Or perhaps the people should admit that copyright, an artificial construct which is not rooted in natural property, was inherently broken and is not (or at least no longer is) a net benefit to the society and simply adapt around it.
Only if I as a human being get the same rights to e.g. Microsofts code as they get to mine via legal trickery in your world.
You do. Just point the LLM towards windows.exe and tell it to party on.
Won't work this year, but it probably will next year. Copyright is done.
My one big issue there is that "adapt around it" tends to look like cryptic black box "Software as a Service" because that gives you a solid moat against competition / reproductions. I would like to be able to pay for software whose function is wholly transparent to me.
That said, I mostly end up using open source for the same reason
Is that so? That's just one side of the push and pull.
Feed a bunch of SaaS screenshots to LLM and you get a GTK+ app for yourself.
I was watching a video talking about how a world before copyright allowed innovation to spread quickly and allowed people miles away to iterate faster. Even if LLMs reproducing feature perfect software is deemed a copyright violation, people will just do it privately and use the software themselves.
I’m personally waiting for LLMs to get so good that I can make music and movies based on my favorite ones. I probably could never release it to the public, but being able to make it and enjoy it myself would be amazing.
Another thing that can help contextualize this phenomenon is mix tapes, which are fair use in the US.
Modifying/modding/remixing software was simply not as feasible as music, but LLMs made it possible.
There is something similar happening in the game modding communities. One of my favorite streamers had claude write a little mod to change the UI of KSP so it kept with the larger fanciful theme of the game, over being so sci-fi-ish. He didn't like, he changed it, he's probably not going to release it because of sensitivity in the broader gaming ecosystem.
I have agents maintaining several patches to my main tools, not forking, not sharing (some have no interest), personal adjustments
2 replies →
That world worked because the ones doing the research were either self-sufficient hermits (often self-sufficient by necessity as they were outcast for "being mad"), financed by the Church or financed by a rich person (usually the fiefdom's ruler, sometimes independent wealth).
Copyright, patents and IP are the evolution of our (Western) way of converting research into a form of financial investment.
1 reply →
> Right now, the settled law is that such an LLM reproduction is 100% legal.
Where did you hear that? Because it is 100% untrue and is the opposite of current legal guidance from reputable legal expert
There was a major lawsuit about this involving Google vs Java years ago. Copying code might not be legal, but an LLM can use the application, learn how it works, write unit tests around that, and then fill in code that passes those unit tests.
The current fun loophole is to have the LLM decompile the existing code, build unit tests around that, and then have a different LLM model build the code that satisfies those tests.
Either way, you can absolutely get a "clean room" result from an LLM.
Yeah a “properly conducted clean room reimplementation” is legal, not an “LLM reproduction” full stop.
And there’s a huge foot gun here in that many LLMs have been trained on copyrighted code that may include the subject of your reimplementation in some cases, which pose a risk of breaking the clean room.
> the settled law is that such an LLM reproduction is 100% legal.
My understanding was that re-invention without copying any code is legal. But scraping code from the browser and re-using it is not. I'd love to know how that plays in the courts with LLMs, as their entire model comes from copying code as training material, not writing new code from scratch.
Sorry to have been unclear - I did mean "clean room". But an LLM can easily jump through the loopholes currently required for that.
Regular humans train on copying code too (Stack Overflow, etc.) so unless they were trained on that specific codebase, I really doubt you have any sort of legal standing. And given how little compensation the authors got when their work got trained on, I wouldn't hold out hope for a big payout even then...