Derivative works are generative AI’s poison pill

Meta’s recent Llama 2 launch demonstrated the explosion in interest in open source large language models (LLMs), and the launch was heralded as being the first open source LLM from Big Tech with a commercial license.

In all the excitement, it’s easy to forget the real cloud of uncertainty over legal issues like IP (intellectual property) ownership and copyright in the generative AI space. Generally, people are jumping in under the assumption that regulatory risk is something that the companies creating LLMs need to worry about.

It’s a dangerous assumption without considering generative AI’s poison pill: derivatives.

While “derivative works” have specific legal treatment under copyright law, there are few precedents for laws or regulations addressing data derivatives, which are, thanks to open source LLMs, about to get a lot more prevalent.

When a software program generates output data based on input data, which output data is a derivative of the input data? All of it? Some of it? None of it?

An upstream problem, like a poison pill, spreads contagion down the derivative chain, expanding the scope of any claim as we get closer to real legal challenges over IP in LLMs.

Uncertainty about the legal treatment of data derivatives has been the status quo in software.

Why do LLMs change the game? It’s a perfect storm of three forces:

•   Centralization. Not until the advent of LLMs could a single piece of software generate variable outputs that were applicable in endless ways. LLMs produce not just text and images, but also code, audio, video, and pure data. Within a couple of years, long before the case law on IP ownership and copyright around LLMs settles, LLM use will be ubiquitous, increasing exposure if risk were to flow past LLM vendors to LLM users. This applies not just to copyright-related risk, but also to risk related to other possible harms caused by hallucinations, bias, and so on.
•   Incentives. Copyright holders have an incentive to argue for the broadest possible definition of LLM derivatives, as it increases the scope over which they can claim damages. Perversely, so do the major platform companies when imposing license restrictions in their total warfare with other platforms. The Llama 2 license is a case in point: section 1.b.v prevents using Llama to “improve” non-Llama LLMs. Fuzzy definitions benefit rights holders and whoever has the biggest legal war chest.
•   Risk-shifting. Software platform companies are masters at shifting risk to their users. The software running the world today comes with an (extremely) limited liability license. Make no mistake: The major platform companies developing LLMs will try to shift risk to their users through legal agreements as well as political means. It’s one of the reasons Big Tech urges AI regulation: Think about how Section 230 protects social media platforms, despite the editorial-like role of algorithmic amplification.

If the courts rule that companies that train their models on copyrighted material are infringing on copyright, there are two distinct types of risk the enterprises that have built on top of those models will have to address:

• Platform risk. Will the vendor pull the model off the market? If so, will a replacement model with comparable functionality be available? What will be the total effort of retuning models and prompts? How long will it take?
• Pricing risk. If the vendor does not pull the model off the market, will the cost of using the model change due to the need to make copyright payments or introduce additional costs in developing or operating the LLM?

Of course, LLM vendors will argue that models themselves are not infringing, even if trained on copyrighted material. Models are just data that looks nothing like the source material. It is model outputs that may infringe on copyright (e.g., consider the prompt “Reword the lyrics of Blinding Lights by The Weeknd.” ChatGPT’s answer was this).

If the courts agree, enterprises have to manage another risk:

• Flow-down risk: How does an enterprise ensure that its use of an LLM doesn’t violate copyright? How far does the risk extend beyond the direct outputs of the LLM to their derivatives, the value created by people, software and systems using those outputs?

Understanding the risks posed by generative AI’s poison pill also gives enterprise technology leaders the tools to manage them.

Our advice:

• When considering LLM licenses, aim for clear ownership of LLM outputs and derivatives, and unrestricted use for improving other LLMs. In the absence of a clear definition of an LLM output derivative, establish a thoughtful policy about what is the copyright equivalent of transformative change of the LLM outputs. (Lowercasing the output probably isn’t, but summarizing the output using a different LLM probably is.) This will act as a firewall against flow-down risk.
• When considering paid licenses, demand insulation from certain kinds of risk and address the economics of the relationship, should risk flow through the vendor to your business in the future. It is a lot cheaper for a large LLM platform vendor to buy IP use rights important to your domain; or, failing that, set up specific types of insurance than it is for their customers to do it. There’s ample precedent in the cybersecurity space, with some vendors bundling ransomware insurance. In generative AI, Adobe is offering full indemnification for ‌content created through Firefly, and Writer offers full indemnification for content generated through its platform.
• Don’t ignore the political side: If LLM users do nothing, the end outcome will be meaningful regulatory protections for the large LLM platforms and Big Tech at the expense of LLM startups and users. ChatGPT Plus and Microsoft’s expected pricing for generative AI capabilities in Office fall in the $25–$30/month/user. At that level of revenue, most types of risk shouldn’t flow down to paid users.

The world of software had a similar issue with “viral” / “copyleft” open source licenses focused on derivatives, epitomized by the GPL. Open source exploded at the same time as SaaS and cloud computing did. For better or worse, SaaS applications and cloud infrastructure got around the GPL poison pill by not distributing software. The AGPL license closed the loophole and is often the choice of open source efforts backed by businesses that want to exert control over their value chain (e.g., MongoDB, Nextcloud, OpenERP, and RStudio).

By contrast, most organic open source projects use more permissive licenses (Apache 2.0, BSD, MIT). Will open source LLMs save the day? They might help enterprises get around certain commercial LLM license restrictions, but they don’t insulate LLM users from copyright risk.

Just as the world of open source licensing bifurcated, so will the world of LLM vendors. Some platforms will follow the status quo of “push all risk to users.” Other enterprise platforms will differentiate by partnering with their customers to manage risk. Risk management will take many forms, from verticalized training over clearly defined input data with traceable usage rights all the way to services that, similar to certain private messaging platforms, make the enforcement of any legal action against their users impractical.

Balancing LLM capabilities with risk management is likely to get more complex as we ease out of the Wild West era of AI — but certainly well worth the effort.