The recent release of AI models by technology giants Google and Meta has highlighted an alarming trend in the licensing of these technologies. While promoted as “open,” these models come with restrictive licenses that pose significant legal and practical barriers to companies that desire to integrate them into their products or services.
Google’s Gemma 3, for instance, is a line of AI models that have been praised for their effectiveness but criticized on the terms of their licensing. The license permits Google to restrict use if it discovers that the model is being deployed in violation of its policies or the applicable legislation. This is a source of uncertainty and risk to companies that would like to deploy them for business purposes since they fear surprise enforcement of these terms.
In the same vein, Meta’s Llama models are subject to unique licenses that restrict how developers can apply the output or results of such models. In particular, developers are not allowed to use Llama 3’s output for improving any model except Llama 3 or its variants. Further, corporations with more than 700 million monthly active users need to acquire a special license to deploy Llama models, adding to their complexity of adoption.
These restrictive terms do not apply to Gemma and Llama alone. Proprietary licenses are more often opted for by most AI model developers than industry standards like Apache or MIT. While this can be utilized for furthering some causes, like limiting commercial use towards advancing scientific research, it proves challenging to enable businesses to navigate the law. Small companies are discouraged by the potential of legal as well as financial expenses related to these custom-made licenses.
These licenses, according to experts, cannot be considered “open source” due to their restrictive nature. Florian Brand, a research assistant at the German Research Center for Artificial Intelligence, says that special licenses are not like normal, approved licenses, causing problems and expense for businesses, especially those without in-house legal personnel. Simply by existing, even without being enforced, the risk of complicated legal language repels use, forcing companies to go toward models with simple, standard licenses.
The impacts of such restrictions percolate well outside of commercial usage. They find resonance also within AI researchers because the uncertainty and possible legal stakes of such models limit their deployment within bigger research projects. These licensing practices pervade the total AI environment and influence the general AI ecosystem through stifling collaboration and innovation.
There is a call to the AI market to synchronize to existing open-source principles to provide a truly open environment. This would be a move towards more liberal licenses that allow free integration, adjustment, and distribution of AI models without fear of sudden changes in the license or juridical ambiguity. As per Yacine Jernite, a machine learning and society lead at AI startup Hugging Face, easing these conditions would lead to greater use and success, as in the case of models that have achieved widespread dissemination despite restrictive licenses.
Despite these problems, some models, including Meta’s Llama, have been widely used, with hundreds of millions of downloads and applications in big companies’ products. Their potential would be even greater if they were licensed under more open frameworks. The landscape of AI model licensing is one of uncertainty, restrictive terms, and misleading claims of openness, which highlights the need for clearer standards and collaboration between providers and users.
