By now, most people will have heard about tools like ChatGPT and Stable Diffusion – AI tools that can generate complex text and images.
They are types of “generative AI”. Unsurprisingly, their release has prompted a raft of litigation based on breaches of rights asserted by content creators. Most of those challenges allege the output appears to be a copy of their work, or occurred because their work was used in training the AI.
There are three major lawsuits targeting AI tools and some have potential implications for New Zealand. As far as we can find, no defendants have filed full responses to the various complaints.
GitHub Copilot
GitHub is a Microsoft-owned code repository, meaning people go there to develop and publish software code. Once published, that code is often subject to licence terms dictating how it may be used. In the not uncommon circumstance that the code is not subject to a licence, GitHub’s own materials acknowledge that copyright law does not allow reproduction, distribution or the creation of derivative works.
Open AI and Microsoft collaborated to make a tool called Copilot. Copilot helps people write code, a little like autocomplete on your iPhone. The language model for Copilot was at least partly trained on the GitHub code repository.
In November 2022, lawyers filed a class action lawsuit against Microsoft, GitHub and OpenAI, asserting that:
- Microsoft and OpenAI trained Codex (the language model used to power Copilot) and Copilot using code published on GitHub;
- doing so violated the open-source licences under which that code had been published; and
- Codex and Copilot output copyrighted materials without following the terms of the relevant licences.
They claim relief on the basis of breaches of copyright legislation, breach of contract, unjust enrichment, privacy and personal data breaches and a range of other causes of action.
Based on previous publications, it is anticipated the defendants will argue that their generative AI system has learnt things by looking at what’s publicly available online and simply use what they learn to create new works (like a human does).
They are likely to say that falls within the US “fair use” doctrine, allowing limited use of copyright material. It remains to be seen what defences or arguments may be made in connection with the other causes of action.
Artists vs Stable Diffusion
Stable Diffusion is an online tool that will generate images from text prompts. It is owned by Stability AI.
In January 2023 the same group of lawyers involved in the Copilot litigation filed a second class action on behalf of three artists against Stability AI, DeviantArt and Midjourney. The lawsuit makes similar claims to the Copilot litigation (breach of contract, violation of the DMCA and a few others), and also alleges direct copyright infringement.
Stable Diffusion and DreamUp [the DeviantArt generator] was trained on a public dataset of 400 million images and captions. The lawsuit claims that training the tool in that way infringed artists’ copyright, and enabling the reproduction and distribution of derivative works based on those images is also copyright infringement.
Getty Images vs Stable Diffusion
Getty Images, a stock photography company, has also sued Stability AI, both in the UK and in the US. The claim is that Stability used copies of Getty’s images to train its AI model and at times produces images that are highly similar and derivative of the Getty Images library and frequently include something that looks like the Getty Images watermark.
In January, Getty Images announced it commenced proceedings against Stability AI in the UK on the basis of copyright and “other intellectual property rights”. The UK has a “fair dealing” exception to copyright law (narrower than the US “fair use” doctrine and more similar to our own), and has specific exemptions from copyright for “text and data mining” for noncommercial research.
In February Getty Images filed a complaint in the US making claims for breach of copyright, trademark and unfair competition laws.
The issues
These lawsuits centre around one main issue: to what degree is it permissible for a software tool to look at things it finds online, learn from them, and make that derived knowledge available to other users?
The two strongest legal challenges seem to be in contract law (“you trained your model on data obtained through breach of contract”) and copyright (“training your model on my art is a breach of copyright”). The copyright claim, in particular, is likely to turn heavily on the American concept of “fair use” and whether that provides a defence, and the UK’s text and data mining exemption.
There are already some lessons that can be learnt: New Zealand developers of generative AI tools need to be cautious about what data they use to train their models. That is the case even if these cases fail in America. New Zealand lacks the broad “fair use” defence to copyright infringement, so a claim in copyright infringement here might go further than it does in America.
And for those excited by the idea that there might be a ChatGPT-style AI that can produce basic advice and replace law graduates, look out for the accuracy ratings of these tools and perhaps consider how that phone call to the liability insurer might go – in the short term at least.
Andrew Easterbrook is a director at Webb Ross McNab Kilpatrick Lawyers (WRMK Lawyers).
0 Comments