Flammie A Pirinen on github pages
TLDR: I think that all current open source licences (also default licence for unlicenced data) already say: you can use this data to train an LLM, but if it’s e.g. used in a text generation to re-generate more or less verbatim copies of this data, someone needs to follow this licence terms (e.g. attribution, share alike, non-commercial). This is not done in 2020–2025 so far. We shouldn’t need a new licence to force creators or users of contemporary AI to follow the laws ad regulations.
The large language model scene has brought more focus about intellectual property rights, licences and all that legal stuff, about texts in general. While I generally am not a big fan of dealing with legal things, this is a highly relevant question for most of computational linguistics that I work with. There are few separate issues here overall and I try to discuss about all of them, but the text may become a bit confusing as a result.
As a sidenote: I use CC BY-SA licences for most of my texts and GNU GPL licences for most of my software, if you include them in your LLM generative AI training set, and your generative AI generates these texts or source codes, they need to include attribution and the resulting new text or programme needs to be shared alike with the same licence. Otherwise the usage is against the law and the licence, whether it’s on creator of the generator or user of it that is responsible, I am not sure, but the legality and the morality of the issue is quite clear. Using LLMs to launder licence terms and authorship away is absolutely immoral and most likely illegal too.
Only data that is totally and completely free of copyrights can be used and re-used without attribution. This is data that whose authors have died long ago, or that has been very explicitly licenced so, and usually data produced by governments such as law texts. And even for these texts, it is just good manners to mention to original authors.
For everything else, the legal requirement of any re-use, such as generating verbatim copies of texts via generative AI, require attribution. This is not being done currently. We can debate whether it’s the responsibility of the generator bots authors or the user of the generator bot, but someone is legally required to do it. This is explicitly mentioned in most permissive open source licences, such as CC BY or GNU GPL, it is also implicitly the licence of all the texts where licence has not been mentioned—there is no publically available data that is free for grabs like OpenAI and other big corporations keep saying.
One common argument that there is is based on the USA law of fair use, this is specific to USA, applies to borrowing of few sentences instead of paragraphs and chapters, and does still explicitly require the attribution. This seems to me largely irrelevnt for the discussion, the intention of fair use is quite far from storing whole texts in a machine that can pretty much copy-paste large snippets of it verbatim. AFAIK there’s no commonly accepted fair use outside USA and when companies like to point out that similar rules exist in e.g. Europe it quite often turns out that they are meaning rules relevant for using text snippets in very restricted cases like teaching material, reviews, parodies, etc. And all require attribution.
Some companies that do not call stealing data fair use simply keep using the phrase publically available when rationalising that everything that can be scraped off the internet is free to be used in whatever way possible. This is even more absurd, because as a starting point, copyright applies to all written text from the moment it is created. In fact you need to be very explicit to give away some rights to text you have written, so claiming that anything you can get your hands on is yours is just bonkers. And to top that off, most texts that are scrapable probably do have even explicit and detailed copyright and licencing terms, and it’s the scrapers responsibility to find this out.
See also:
This is an interesting question that not many have talked about. My starting point is that freely licenced texts as well as texts that don’t have explicit licence (so-called publically available data or web crawls), already have a copyright that requires re-use be at least attributed at all times. So whose responsibility is it to provide attributions.
There’s a lot of misuse of terminology in the AI and LLM world, lead by the absurdly named OpenAI, a highly closed source for-profit corporation trade secrets and whatnots. To counter that I think it was meta that started selling “open source” AI, which in fact is kind of a downloadable binary that contains some parametres that others can use to run it, and perhaps even mod and finetune, a software equivalent would be free software perhaps. Open source, as we know it, either by Free Software Foundation or other open source movements, requires access to all necessary data to re-create the model from the scratch, and edit and remix it, that’s why it’s called open source, you get the sources for re-building. In LLM context this clearly means all the data that was fed to the LLM, there’s no ifs and buts about it. That is the only way to study the products and remix it truly to ensure it doesn’t contain anything unwanted.
See also: