Who owns the output of LLMs?
IP in AI outputs
Copyright exists to incentivise, reward and protect artistic creativity. When a human writes a play or paints a picture, they can expect to rely on copyrights to prevent others from plagiarising their work. Does it make a difference if a human makes use of AI as part of the creative process?
The businesses behind AIs don’t shed much light on this issue. OpenAI’s T&Cs, for instance, clearly say they assign to the user ‘all [OpenAI’s] right, title and interest in and to Output.’ But they don’t promise that there is actually any right, title and interest – just that if they’ve got it, you can have it. This might turn out to be a big, fat nothingburger. But they won’t stand in your way: ‘you can use Content for any purpose, including commercial purposes such as sale or publication’. Nice.
Like any good legal question, the actual answer is ‘it depends’. It depends on the extent of AI involvement, and this will lie somewhere on a spectrum from 100% human generated to 100% AI generated.
100% human generated
If you’re just using an LLM as a tool in the same way you might use Photoshop, or a paintbrush and easel, there’s no problem. The work is going to be your original effort (assuming you’ve not copied it from somewhere else) and so it’s going to be copyright. All your copyright. OpenAI won’t try to claim it, any more than Adobe or Winsor & Newton could. They just supplied the tools, the actual act of creation was entirely down to you.
100% AI generated
On the other hand, if a work is produced entirely by an LLM — you prompt, it outputs, the end — most laws are clear that no copyright will exist.
In the US, the law says a copyright work must be ‘original’ and that this means it has been independently created by an author and displays a minimal level of creativity. Even if you could argue that LLMs are capable of demonstrating creativity (and we’re certainly in that camp), there’s no way to overcome the requirement that the author must be a human. As well as SCOTUS, this is also the considered position of the US Copyright Office, which won’t accept copyright registrations from, variously: monkeys, aliens or AIs.
Australia takes a similar approach – the High Court has held that copyright works must have some ‘independent intellectual effort’ and, once again, this means a human author. AIs and other computers can assemble information, but are not deemed capable of expressing it in a copyrightable way. And copyright is all about individual expression of ideas and information. Similarly, in the UK and EU the law says that works can only be ‘original’ if they are ‘expression of the intellectual creation of [their] author’.
There is a potential exception, but it’s probably more apparent than real: in some countries such as the UK, Eire and New Zealand, the copyright laws specifically refer to ‘computer-generated works’ ie works without a human author. But copyright only applies when these computer-generated works are ‘original’ and so we’re back at square one again: no human author, no originality. It’s a loop back to a dead-end.
Anywhere in the middle
This is where it gets tricky, and is unfortunately also where most works generated with AI are going to fall: you prompt, it outputs, you edit and process/tidy up/finesse/filter…the end.
Something around 90-99% human-generated is probably going to be fine. Likewise, 90-99% AI-generated is likely going to be a no-no. But it’s the middle ground we’re interested in and it’s anyone’s guess where the lines should be drawn. The US Copyright Office says that where an author ‘elaborates or re-elaborates’ a work produced with the assistance of an AI, copyright will only apply to the bits the human author is responsible for, and then only if those bits constitute an original work. This obviously only applies directly in the USA but it’s probably not a bad indication of how things will pan out in the rest of the world too.
Look, for a few reasons it’s probably best not to rely on AI to produce the whole or the bulk of anything you’re going to be too concerned about (eg because it’s important or valuable etc).
First, LLMs are essentially ‘cliché engines’ – they tend to produce output which is safe, vanilla, middle-of-the-road, standard. And if they do it for you, they’ll do the same for the next guy too. So you may both end up with the same or similar starting material.
Second, you don’t know where that material has come from. It might be someone else’s copyright material which has just been scooped up and served out to you. There’s an infringement case on at moment arguing just this, and some of the images in question have been produced with copies of third-party watermarks included, which kind of gives the game away!
Third, even if the material hasn’t been plagiarised, you really can’t rely on there being any copyright in that output, only in what you do with it. Think of the AI output as a lump of clay and it’s up to you to turn it into an original work which the law will recognise and protect. The closer you leave your work to the raw lump of clay, the less you can claim it as ‘yours’. But work it into something unique and special, and now we’re talking.
Fourth, decision makers tend to be biased against AI. There’s no rhyme nor reason for it, but there it is. It’s therefore probably worth trying to downplay the involvement of AI in anything you produce. We’re not suggesting you deny all involvement, not should you lie or dissemble (especially not where you’ve got to reference source materials). But going around shouting about how clever you are for making something with AI is not going to get you very far. Best to treat LLMs like just one of several tools in your creative toolbox.
This may all change in the next few years as lawmakers around the world get into gear (some countries are looking at introducing entirely new IP rights specifically for AI-generated works). But for now. that’s all in the future.
Hope this helps.