8 ms·
This is pretty interesting, and I have conflicted feelings about the (seemingly obvious) outcome of this trial. I wonder, if MS and OpenAI win, does that mean
by bityard 2y ago
This is pretty interesting, and I have conflicted feelings about the (seemingly obvious) outcome of this trial.
I wonder, if MS and OpenAI win, does that mean it will be legal for anyone to take the leaked source code for a proprietary product, train an LLM on it, and then ask the LLM to emit a version of it that is different enough to avoid copyright infringement?
That would be quite the double-edged sword for proprietary software companies.
- ChrisMarshallNY 2y agoI suspect that this is exactly what will happen; not just with code, but also prose and artwork. Someone is likely to design an LLM that is specifically trained to do exactly that. Lots of money to be made...
- devmor 2y agoOn the matter of artwork there's no need for suspicion - it is and has been happening for a while now. There are entire online databases dedicated to providing non-consenting artist's "styles" as downloadable model parameters by name.
- ChrisMarshallNY 2y agoTry getting Mickey Mouse comics. That should be fun...
- satvikpendem 2y agoStyle is not copyrightable so I see nothing wrong with making essentially a robot that can paint in the style of someone else.
- falcolas 2y agoIn isolation, no. But the produced works can be too close for fair use (as demonstrated with the Prince pieces by Andy Warhol), and passing it off as a piece from the original artist can open you up to forgery/fraud charges. To put another way, the motivations to produce art in another artist's style can still land the artist/buyer in legal trouble regardless of fair use.
- satvikpendem 2y agoYes that is true, but I don't think the people who use style transfer are actually passing it off as the original, they just like it for the aesthetic value of their own images. In other words, no one using the Van Gogh LoRA is actually trying to forge the Starry Night.
- falcolas 2y agoGiven the value of an "authentic" painting of the Starry Night (or more realistically the value of something forged in, say, Samwise Didier's style) I can't agree with "no one". I have to imagine that it's likely quite popular to sell AI generated art that mimics or copies existing works.
- satvikpendem 2y agoDo you use AI art generators? Flaws are extremely easily found out, it is only good for a rough snapshot (without much fiddling and even then, artifacts remain). I can guarantee you it is definitely not popular to sell existing works made with AI, you are better off hiring an actual forger. In fact, your suggestion is even the first I've even heard of such an idea.
- 8organicbits 2y agoI guess there's always a greater fool, but forging an oil painting using AI digital images seems pretty far fetched.
- Aeolun 2y agoYou can paint over the printed image? Not that it’d look anything like the artist you are copying, but it’s a fun idea.
- devmor 2y agoThe legality of using someone’s copyrighted work to train a model to reproduce it without their consent is still under debate - but the morality of the act at least, is not related to its legality - be it positively or negatively; and I personally consider it abhorrent.
- satvikpendem 2y agoUnder what morals do you consider it "abhorrent?" I bet got a straight answer from those I've asked about this as the counter arguments seem too easy to make.
- devmor 2y agoIt's just pure exploitation. You're using the product of someone's work to create a machine that takes away their work.
- struant 2y agoWhy is doing a task with a machine suddenly objectionable when the same task performed by humans is perfectly fine?
- sensanaty 2y agoA man with a small canoe catching a few fish with a fishing rod for his dinner is very different to a commercial fishing vessel trawling through the ocean with a massive net to catch thousands of fish at once. The two are treated differently under the law, and have different rules that apply to them due to the difference in scale. Scale matters, and the scale that computers/these AIs operate under are absurd compared to a person doing it manually.
- satvikpendem 2y agoWhy does scale matter in terms of AI? Just because a computer can do it at scale doesn't mean it should be treated similarly to your analogy. Rather than using an analogy, please tell me why it matters that computers can do something like AI at scale rather than individuals doing it.
- CuriouslyC 2y agoI sure wish I could non-consent to people observing me in the world, I'd like to move through society invisibly and only show myself when it benefitted me. Unfortunately, the only answer is to stay inside if I don't want people to see me.
- vkou 2y ago> I sure wish I could non-consent to people observing me in the world, You aren't allowed to use photos featuring a non-consenting person to, for instance promote a product. You are allowed to use photos including a non-consenting person. There's a lot of complicated law, differing between different jurisdictions to cover this question, and to balance the needs of the public with commercial desires. It's not as simple as you make it sound, and there's no reason we should just default to bending over backwards for commercial interests. Laws exist to serve society, not the other way around.
- CuriouslyC 2y agoI'm sure that the people who are being constantly victimized by paparazzi would like to know those rules that you just quoted, and have them be enforced.
- vkou 2y agoIf you had done a little research into this question, you'd realize that 1A use cases ('journalism') are treated by law quite differently than use of likeness for commercial intent. This is my whole point. There isn't a single, one-size-fits-all rule that a five year old can comprehend that describes any particular country's legal framework around the many, many different dimensions of tension between public and private interests on this incredibly broad question. And none of the existing frameworks fit the new use cases well, and we should probably have an open political debate about what we want to do going forward.
- CuriouslyC 2y agoI'll happily take your picture against your will and put it on the internet with the tag "vkou mad at photographer, news at 11"
- ADeerAppeared 2y ago> Someone is likely to design an LLM that is specifically trained to do exactly that. Perplexity AI.
- chimeracoder 2y ago> Perplexity AI. How does this describe Perplexity AI more than any other LLM?
- ADeerAppeared 2y agoI am referring to their service rather than their LLM in specific. Perplexity is in the business of using an LLM to paraphrase existing content, then serving that up as their own "work" in a way that directly harms the original content they took. It's not even a question of "Is AI training copyright infringement", they're just doing copyright infringement with AI. And it's horribly common already.
- mcmcmc 2y agoThey plagiarize and blame it on the third party service they use for web scraping https://www.theverge.com/2024/6/27/24187405/perplexity-ai-twitter-lie-plagiarism https://www.theverge.com/2024/6/27/24187405/perplexity-ai-tw...
- epolanski 2y agoI feel like what really matters is who has more money to throw in tribunals. Somehow I feel if it was "Adobe vs dev that claims his code was spit by copilot" it would not end the same.
- crote 2y agoI was mainly inspired by this section: > Specifically, the judge cited the study's observation that Copilot reportedly "rarely emits memorized code in benign situations, and most memorization occurs only when the model has been prompted with long code excerpts that are very similar to the training data." That almost sounds like it'd be fine to train an "art transformation model" which takes an image and transforms it, which for all the frames of a specific Disney movie just so happen to output the very next frame...
- saint_fiasco 2y agoThat sounds like the opposite from the quote. The art transformation model you propose WOULD emit memorized art in benign situations, so in that judge's opinion it WOULD count as plagiarism.
- TomatoCo 2y agoWith how modern video codecs use data from previous frames you could make a not-entirely-specious argument that we already have a tool that can do this and it's called ffmpeg.
- anticensor 2y ago> "art transformation model" There is a reason a famous AI model architecture is called transformer, it is pretty much optimised to be good at transforming artistic and intellectual works.
- devmor 2y agoFollowing existing law and applying reasonable expectations, I would point to the old adage "intent is 9/10ths of the law". It would probably be legal to do this, as long as no one could reasonably show that you intentionally trained the LLM on said leaked source code with the intent to reproduce the product. Of course, civil suits could be another matter entirely. If you pick a product to rip off that's owned by a multi-billion dollar company, all that can save you is the ethical limits of their legal team's consciences.
- spencerflem 2y agoNot unless a big company is the one doing it lol
- pennomi 2y agoOr even those AI-powered decompilers people are working on… you could clone virtually any software with that. Surely there will be limitations.
- beeboobaa3 2y agoThe limitation is the amount of money & political power the owner of the software you're cloning has.
- wongarsu 2y agoThe source code of Windows XP is widely available. Same with a ~2 year old version of Bing, Bing Maps, Cortana etc. Yet that doesn't seem to have had major negative effects on those products. If anything having the Windows source code available seems to be a net boon for Windows development. Sometimes looking at the source is just better if the documentation is unclear.
- userbinator 2y agoMS probably hates that the source for XP/2K3 leaked because it means more people will put in effort to fix and extend/backport, even if it's not truly legal, when MS would rather coerce them into the latest most invasive and user-hostile version. Also because projects like NTVDMx64 show how some of their decisions have been political instead of technical as they like to claim. Far less people care about Bing or Cortana.
- mr_toad 2y agoIf you compiled it and the resulting binary was substantially similar to the original you’d likely get sued.
- Legend2440 2y agoI mean you can legally do this by hand right now. That's how they cloned the IBM bios back in the day. IBM sued and lost.
- marcosdumay 2y agoNo, that's not. They cloned the bios by observing how it behaved and writing code that behaved the same way. Nobody even looked at the bios code.
- wvenable 2y agoThat's not how they did it. They had one team read the BIOS source listings in the IBM PC Technical Reference Manual and create a technical specification and a second team take that specification and write a new BIOS [1]. The second team never saw the original code so therefore they could not have copied it. To do something similar with AI, you really need to train one AI on the source code and then have it explain that code to a second AI that never saw the original code. [1] https://en.wikipedia.org/wiki/Phoenix_Technologies https://en.wikipedia.org/wiki/Phoenix_Technologies
- axus 2y agoI thought there was a "clean room", where the people reading it and the people writing it were different; and they made a written specification instead of a Vulcan mind meld.
- jeroenhd 2y agoA Wine fork built using an LLM trained on leaked Windows code might be pretty useful.
- Spivak 2y agoNo, because judges aren't robots applying the law like code. Intent matters. If you do this it will be painfully obvious that your intent is to duplicate a large body copywritten code.
- yellowapple 2y agoIt's painfully obvious that the intent of GitHub Copilot is to duplicate a large body of copyrighted code.
- tpmoney 2y agoIt doesn't appear to be painfully obvious. Both because they're not losing court cases yet, and there's a huge swath of non copyrighted code being produced by co-pilot every day. By contrast the plaintiffs apparently were unable to induce Copilot to duplicate any parts of their code.
- Bognar 2y agoOh so that's why Copilot has a filter to prevent suggesting copyrighted code, because the intent is to duplicate copyrighted code. It all makes sense now.
- yellowapple 2y agoThat's exactly what I expect to happen with the source code to Microsoft's own software products, namely Windows. Hilarity will ensue :)
- mr_toad 2y agoThe misappropriation of the code (a trade secret) would likely be grounds for legal action against the people who stole it and the people who received it. A lot depends on jurisdiction. But if it was made public and then if an unrelated third party were to re-write the code in such a way that it was non-infringing, then it would be non-infringing. That’s just a tautology.
- throwaway562if1 2y agoLet's be honest: It will be legal if you're a $3 trillion company, and not if you're not.
- stale2002 2y agoBy definition you are allowed to take leaked source coded and change it enough such that it avoids infringement, and this will avoid infringement. The LLM has nothing to do with it, and isn't required here.
- elzbardico 2y agoYeah. In a ideal world where an open source developer gets equal treatment from the law facing a giant corporation with hordes of very expensive lawyers and "technical experts".