Comment by kmeisthax | Hacker News Reader

Comment by kmeisthax | Hacker News Reader

> That "Vino" driver for DisplayLink driver was a clean-room reverse-engineered DisplayLink driver that was developed in part by AI and written in Rust.

It is my strongly-held opinion that, regardless of what you think about AI in general, reverse engineering projects ought to stay as far away as possible from LLM code gen as possible. They are a very easy and straightforward way for a copyright owner to claim access to their code, since basically ALL code was trained on. Unless you have hard evidence the particular bit of code you want to RE was not in the training set, you should not use an LLM to reverse-engineer or rewrite it.

This is not a moral argument; as much as I hate AI companies playing fast and loose with other people's work, the underlying arguments that make RE legal work much better if you do not look at the source code. And while US judges are straining to keep AI training fair use, that does not make the AI a copyright laundering device. A copyright owner can absolutely reach through a fair use to get to you. Think of it like if you downloaded 10 different YouTube video essays on the same movie and attempted to stitch them together into a normal copy of the underlying movie. Obviously, your use has a different (and worse) legal status than the reviews do.

Also, fuck Synaptics.

I’ve been warning about this for a long time.

With typical clean room RE you get two artificers with clean copyrightbility - a fair use spec describing the system, a new implementation. This is not legally required, but the paper trail is a good defense.

If AI writes the spec, it may have just tainted itself if that work is then used to in future training. The output is not protected by copyright, but that probably doesn’t matter. If AI writes the output, there is no protection over that either, which may or may not matter.

Ultimately, however, this is a mechanical, derivative work. It isn’t washed because no creative human expression was introduced in either phase.

If the phases are split into AI writing the spec and human writing the implementation, and the spec is determined to be infringing, the implementation is infringing. The AI spec has no judgement over what’s copyrightable expression in the original, and there’s a high likelihood it produces an unauthorized derivative.

If a human writes the spec and an AI writes the code, the output is probably clean, but also not copyrightable. This is less of a concern for open source projects written by a single maintainer, but may have effects for larger projects or proprietary projects. However, writing the spec is the harder of the two parts and what people like using AI for.

Except in the most egregious cases, AI is not going to be found to be derivative or infringing. The consequences go far beyond this particular type of situation; it is being used everywhere by the richest corporations to produce all kinds of content. They will not allow those works to suddenly have questionable provenance.

> They will not allow those works to suddenly have questionable provenance.

Even though many of those "richest corporations" trained their models on huge quantities of torrented ("pirated") copyrighted content and other such "stolen" data. Nothing at all questionable there... LOL

Don’t look at tech companies that don’t care about laws and price that into their costs. Pay attention to publishers and what they allow. USCO has written opinions on this and doubled and tripled down on those opinions. AI doesn’t magically erase copyright.

Mostly agree. It's been super helpful for personal use but I don't think it's possible to achieve a true "clean room" implementation. Anything copyright encumbered should stay away.

On the other hand, it's been great for poking at cheap IoT devices and figuring out how to interact without crappy proprietary apps.

Counter Argument, if the code is propietary and well kept, how did the AI got trained on it? Didn't the copyright holder fail to protect his IP by letting AI train on it? If the specification is only released under NDA, how could a AI be trained on it?

I think the copyright holder should prove first in court, that his super secret specs or code was effectively used for training the AI.

But when LLM is used to create an image or video, the user usually consider the LLM as a copyright laundering device and claim full copyright on the produced image or video, regardless of what the licenses of the training set would be. Why coders don't get the privilege of that?

No, you cannot launder copyright through an LLM. It's the same for both images, video, and code; and that question is independent of whether or not it's legal to train the LLM on copyrighted data. The LLM training gets to be fair use, but fair use is not transitive. Infringement reaches through the LLM to get at any outputs which are substantially similar to the original no matter how many fair uses you put in the middle.

> Unless you have hard evidence the particular bit of code you want to RE was not in the training set, you should not use an LLM to reverse-engineer or rewrite it.

Wouldn't that be on Synaptics to prove? And with how protective they seem to be of their IP, I doubt their code is in the training set

Yes, but I'm giving the "defendant's lawyer" opinion. In reality, you can be completely tainted and still do good RE work, but proving that in a court of law will be harder.

And furthermore, the real victory is not in the courts; the real victory is staying out of the courts to begin with. If you win in court, you still lost huge sums of money. And a lot of staying out of court boils down to restricting one's conduct to things a plaintiff's lawyer would counsel against suing for. Hence why most legal opinions in the FOSS space are unusually conservative (while still, occasionally, just being completely off-base).

Reply on news.ycombinator.com