‘Open’ model licenses frequently include troubling limitations
Google Unveils Gemma 3 Models Amid License Concerns
This week, Google introduced a new set of AI models known as Gemma 3, which have quickly received praise for their efficiency. However, developers have expressed concerns about the restrictions tied to the licensing of these models, which they believe make commercial use riskier. This concern is not unique to Google, as companies like Meta also impose specific licenses on their AI models, which can pose legal challenges for businesses.
The Impact of License Restrictions
Many smaller companies are apprehensive about using these models due to the potential for unexpected legal consequences. They worry that tech giants such as Google might enforce their licensing terms aggressively, impacting their business operations. Nick Vidal, head of community at the Open Source Initiative, highlighted how the confusing and restrictive licensing of AI models creates uncertainty for businesses looking to adopt them.
Licensing Terms: What You Need to Know
Developers of open models often choose to issue them under proprietary licenses instead of standard options like Apache or MIT. For instance, AI startup Cohere clearly states that its models are intended solely for scientific use, not for commercial application. In a similar vein, both the Gemma models and Meta’s Llama models have licenses that impose various restrictions on their use.
Key Restrictions on AI Models
Meta’s Llama Licensing: Meta prohibits the use of Llama 3 models to enhance any other models besides Llama 3 itself. Additionally, entities with over 700 million monthly active users must secure an extra license to deploy Llama models.
- Gemma Licensing: While slightly less restrictive, Gemma’s license still allows Google to limit usage if it believes the conditions, including relevant laws and regulations, are being breached.
Challenges for Developers
The restrictions associated with these models apply not only to the original versions but also to any models derived from them. Even models trained using synthetic data generated by Gemma would need to comply with the same licensing terms. Florian Brand, a research assistant at the German Research Center for Artificial Intelligence, argued that licenses like those of Gemma and Llama cannot accurately be described as “open source.” He emphasized that many companies prefer licenses that are widely accepted, like Apache 2.0, due to the potential complications of custom licenses.
The Impact on the AI Ecosystem
The peculiar licensing terms pose challenges not just for businesses but also for researchers in the AI community. Many professionals, including Han-Chung Lee from Moody’s and Eric Tramel from AI startup Gretel, agree that restrictive licenses fundamentally limit how models can be utilized in various commercial applications. The complexities involved in determining licensing for derivatives or fine-tuned models could lead to significant legal ambiguities.
Concerns About Market Manipulation
Tramel pointed out a pressing fear: that companies could release AI models as a tactic to later exert control over the market. This fear stems from the notion that a model could be initially presented as "open," yet later become a minefield of legal restrictions, forcing users to reconsider adopting it.
The Potential for Open Licensing
Despite the restrictive nature of these licenses, some models have still gained popularity. Meta’s Llama, for example, has seen extensive downloads and integration into products from leading companies like Spotify. However, Yacine Jernite from AI startup Hugging Face believes these technologies could thrive even further with more permissive licensing frameworks. He advocates for companies like Google to adopt open license structures and engage with users on mutually agreed-upon terms.
A Call for Clarity and Uniformity
Vidal emphasizes the urgent need for licensing structures that allow companies to integrate, modify, and share AI models without fears of sudden changes or legal uncertainties. He cautions against redefining "open" to serve corporate interests and instead calls for alignment with established open-source principles, advocating for a more transparent and accessible ecosystem in the AI domain.
