Licences for local AI models: are they worth anything?

The core logic behind such products is called the model and is usually described by certain values (parameters) arranged in a certain way (topology), which determines how information is processed. Typically, only the topology is determined by a programmer and the respective parameters are trained.

To avoid misunderstandings, these models do not contain "classical" source code. Running them requires other software, which is usually readily available as open-source libraries. Consequently, this software is not typically considered a particularly valuable part of an AI system.

Closed vs. open source

Many of these models are closed source. No one outside their respective providers has access to their parameters and topology or can modify or run them on their own hardware. Without needing to invoke any IP rights, this creates an effective monopoly for their respective providers over each model, since it is kept a secret from aspiring users, which have to subscribe to the provider's service.

However, increasingly impressive models are becoming available for free download. Anyone can download these models, run them locally and modify them for their own purpose. Such "local AI models" are often referred to as open-source models. However, providers of such models typically assert that their use is permitted only if users comply with certain requirements stipulated in a licence.  

Some models, such as Mistral's famous large language models, are provided under Apache 2.0, a rather liberal open-source licence originally intended for normal software to be applicable. Other local AI models, such as Stability AI's Stable Diffusion-Models, Meta's Llama or Bloom's LLM are all provided together with much more restrictive licences that limit use to only certain allowed purposes (such as non-military, non-commercial or solely research purposes).

But does such a licence actually work as intended and is it enforceable? By analogy with other software licences, particularly open-source software licences, one might be inclined to pre-emptively conclude: Yes!

Not so fast

Licence agreements inherently rely on the protection provided by copyright law to enforce their terms on users. However, copyright protection is conferred to software only if certain conditions are met. Crucially, software (understood to be a sequence of instructions to a computer) must be an original work in that a programmer must have used a certain minimum amount of creative freedom not determined by functionality and technical guardrails. If the shape of the program is fully determined by technical requirements, no copyright protection may be assumed.

AI models are not easily assumed to fulfil this originality requirement. For example, the network topology, though involving design choices such as the number and arrangement of neurons and layers, is often highly constrained by technical requirements. These constraints raise doubts about whether there is enough creative freedom in designing the topology to qualify it for copyright protection. The parameters themselves are then trained by a training program ("machine learning") and are not subject to the developer's choices.

Unlike the implementing code, which can be written with ease thanks to existing open-source libraries, the final AI model – including its topology and trained parameters – represents the valuable core asset because of the substantial effort involved in its creation (training processes that require significant computational resources, time and large datasets).

Compliance with local AI model licences may, therefore, not always be enforceable through the monopolising effect of copyright law.

Navigating the licence labyrinth

If the local AI model made available for download does not fall under the scope of copyright law, the enforceability of licence restrictions may depend solely on whether the user is considered to be in a contractual relationship with the model's provider. This may not always be straightforward. Under Austrian law, for example, the existence of such a contractual relationship would need to be proven. Furthermore, there is very little reason for the user to argue in favour of having entered into an agreement that would bind them and such contractual terms could hardly affect any third parties that gained access to such models via "non-official" channels.

This potential deficiency in the ability to effectively and broadly enforce licence terms for open AI models is a novel challenge brought about by the rise of such models. As demonstrated by the leak of Meta's Llama model, this issue is far from insignificant for providers seeking to control the use of their locally available AI models. Only time will tell if this problem will gain greater significance, especially as local AI models potentially grow in popularity.

author: Alexander Pabst