Since the explosion in generative AI, there has been a rash of “decompilations” of video games and various other software (such as this example from earlier today disputed – see below) that have been published to Github and advertised as “Open Source.” That claim is a lie.
The phrase “Open Source” has a specific meaning (and similarly for “Free Software,” by the way1), and it isn’t merely that the source code is there for you to look at. It means that the copyright holder is explicitly giving you permission to read that source code, modify it, redistribute it, etc. Without that element of permission, the code cannot be “Open Source” even if you can physically read it. At best, it might be “Fair Use” depending on the circumstances, but it’s most likely just a fancy means of copyright infringement.
Remember, copyright is a legal construct, not a technical one. It depends much more on the intent of the human doing the copying than it does on the technical details of what they actually did. If the thing they have is obviously intended to be a copy of something else, it is a Derivative Work no matter what technical means were used to create it. That means the original copyright still attaches to it and the person who made the copy doesn’t get to choose a new license for it, “Open Source” or otherwise.
Why YSK:
You don’t have to like the way copyright law works – I sure don’t! – but you do have to understand it because there’s a lot of misinformation going around right now with people claiming things are “Open Source” when they aren’t and a lot of people are going to get in trouble for it. It also dilutes the public understanding of what actual legitimate Open Source software is, which is a problem in and of itself.
Conflating real Open Source software with proprietary software that’s been ‘pirated with extra steps’ is harmful both for developers of the former, who have their reputations damaged by association, and for users/sharers of the latter, who might be misled into not taking the same precautions that they would if they understood that they were dealing with warez. Just because you might think Big Tech can get away with laundering copyright through LLMs – and even that remains to be seen – doesn’t mean the little guys can.
TL;DR: Proprietary software cannot become Open Source software by any means except (a) the express consent of the copyright holder or (b) the copyright expiring and the work becoming Public Domain. Whatever technological end-run you think you have around this legal fact, no you don’t.
EDIT: dispute over example
In giving that example I was relying on the claim in the linked thread, which comes from this guy on BlueSky. Seems like a lot of people think he’s wrong, so maybe that’s not a good example after all.
However, there are also things like this, and those are examples I feel very confident in citing because (a) they explicitly call them “decompliations,” (b) at least one of them has a LICENSE file that says it’s MIT, and (c) there’s zero chance Nintendo or Rare or anyone else legitimately gave them permission for it.
footnote
1 “Free Software” has essentially the same denotation as “Open Source” – close enough that every “Free Software” license is also “Open Source” and vice-versa – but a different connotation. The term “Free Software” tends to get used by people who wish to emphasize the rights of the end-user, while the term “Open Source” tends to get used by people who wish to emphasize that the software is available to be modified.
I haven’t seen that, but I have seen a lot of open-weight AI models being described as open-source, which they really are not, and that misuse of the term is something that I really don’t like. Open-weight models are basically analogous to getting the binary for a closed-source software package, rather than only having to interact with it via a remote server. That’s not to say that that can’t have value, but it is not the same as having the source, which would be getting the training corpus and procedures that would let one reconstruct the model.
There is the possibility of having an Open Dataset of license-free data plus open-weights, which is probably the closest you can get towards Open Source in that space. There’s currently no way to make the model itself transparent simply because of the principle of how these things work.
Best Practice Example i can offer is Apertus 1.5 -> https://huggingface.co/swiss-ai/collections for models and datasets, https://github.com/orgs/swiss-ai/repositories?q=apertus for training code and similar, https://www.apertus-ai.org/ for the homepage
Agreed, and that’s a whole 'nother problem. “Open Source” as a term is inadequate to apply to AI models because the only part that actually counts as “source code” is just the harness/framework that runs it – even closed-weight models can be (and mostly are, AFAIK) technically “Open Source,” but that does you no good when the important part is the gigantic opaque vector of data it needs to actually do anything.
All that aside, I really want to reply to:
First of all, there’s the example I already cited, which is allegedly a decompilation of Adobe Suite that claims on their main page under the heading “Open Source” that “every line is on GitHub under permissive licenses” (even though they’re lying about that because it isn’t even a real OSI-compatible license, LOL).
Second, there have been a bunch of decompilations of Nintendo 64 games including Mario and Zelda ones, and Nintendo, of all companies, sure as Hell didn’t approve that!
That license suggests these people don’t know what they’re doing, or don’t care. Maybe they think they’re geographically out of reach of Adobe’s lawyers. In any case, it’s hardly striking a blow for open-source software when its provenance and legal status are so dubious. FOSS should be done honestly and with pride.