AI Model Licenses: Understanding Restrictions | Open AI

Google’s Gemma 3 and the Complexities of ‘Open’ AI Licensing

Google recently launched the Gemma 3 family of open AI models, which have been widely recognized for their notable performance and efficiency. However, a number of developers expressed concerns on the X platform regarding the potential risks associated with the models’ licensing terms for commercial applications.

Licensing Challenges Extend Beyond Gemma 3

This issue isn’t isolated to Gemma 3. Other companies, such as Meta, also employ customized, non-standard licensing agreements for their publicly accessible models. These terms introduce legal complexities for businesses seeking to utilize them.

Specifically, smaller companies are apprehensive that Google and similar entities might revoke access or enforce stricter clauses, potentially disrupting their operations.

The Open Source Initiative’s Perspective

Nick Vidal, community head at the Open Source Initiative – a prominent organization dedicated to defining and promoting open source principles – explained to TechCrunch that the ambiguous licensing of so-called “open” AI models is generating considerable uncertainty, especially concerning commercial adoption.

“Despite being marketed as open, the actual terms impose various legal and practical obstacles that discourage businesses from integrating these models into their products and services,” Vidal stated.

Reasons for Proprietary Licensing

Developers of open models have justifications for releasing them under proprietary licenses rather than established standards like Apache and MIT. For instance, AI startup Cohere has explicitly stated its intention to support scientific research, but not commercial endeavors, built upon its models.

Restrictions in Gemma and Llama Licenses

However, the licenses for Gemma and Meta’s Llama models contain specific limitations that could create legal issues for companies.

Meta, for example, prohibits the use of Llama 3 model “outputs” to enhance any model other than Llama 3 or its “derivative works.” Furthermore, organizations exceeding 700 million monthly active users are required to obtain a separate, additional license to deploy Llama models.

While Gemma’s license is generally less restrictive, it grants Google the authority to limit usage of Gemma if it deems the use to be in violation of the company’s prohibited use policy or “applicable laws and regulations.”

Inherited Licensing Requirements

These licensing terms extend beyond the original Llama and Gemma models. Any models derived from Llama or Gemma are also subject to the respective Llama and Gemma licenses, including models trained on data generated by Gemma.

Questioning the Definition of ‘Open Source’

Florian Brand, an AI researcher at Trier University, argues that licenses like those of Gemma and Llama “cannot reasonably be called ‘open source,’” despite claims made by technology company executives.

“Most companies maintain a list of approved licenses, such as Apache 2.0, so any custom license presents significant trouble and expense,” Brand explained to TechCrunch. “Smaller companies lacking legal teams or financial resources for lawyers will likely opt for models with standard licenses.”

The Threat of Enforcement

Brand also noted that, while AI model developers with custom licenses, like Google, haven’t actively enforced their terms yet, the mere possibility of enforcement can deter adoption.

“These restrictions impact the AI ecosystem – even affecting AI researchers like myself,” Brand added.

Concerns from Industry Professionals

Han-Chung Lee, director of machine learning at Moody’s, and Eric Tramel, a staff applied scientist at AI startup Gretel, both agree that custom licenses like those associated with Gemma and Llama render the models “not usable” in many commercial contexts.

“Model-specific licenses include specific exceptions for model derivatives and distillation, raising concerns about potential clawbacks,” Tramel said. “Consider a business specializing in producing model fine-tunes for its customers. What license should a Gemma-data fine-tune of Llama carry? What would be the impact on all of their downstream customers?”

The ‘Trojan Horse’ Scenario

Tramel expressed concern that these models could be a “trojan horse,” where a model foundry releases “open” models, observes the development of business applications, and then attempts to gain control of successful verticals through legal pressure or extortion.

“For example, Gemma 3 appears to be a strong release with the potential for broad impact. However, the market is hesitant to adopt it due to its license structure. Consequently, businesses will likely choose potentially weaker and less reliable Apache 2.0 models.”

Widespread Distribution Despite Restrictions

Despite their restrictive licenses, certain models have achieved significant distribution. Llama, for instance, has been downloaded hundreds of millions of times and integrated into products from major corporations like Spotify.

Call for Permissively Licensed Models

However, Yacine Jernite, head of machine learning and society at AI startup Hugging Face, believes these models could be even more successful with permissively licensed terms. Jernite urged providers like Google to adopt open license frameworks and “collaborate more directly” with users to establish widely accepted standards.

“Given the lack of consensus on these terms and the fact that many of the underlying assumptions haven’t been tested in courts, it primarily serves as a statement of intent from those involved,” Jernite said. “But if certain clauses are interpreted broadly, a lot of valuable work could face legal uncertainty, which is particularly concerning for organizations building successful commercial products.”

The Need for a Truly Open Ecosystem

Vidal emphasized the urgent need for AI models that companies can freely integrate, modify, and share without fearing sudden license changes or legal ambiguity.

“The current AI model licensing landscape is characterized by confusion, restrictive terms, and misleading claims of openness,” Vidal concluded. “Instead of redefining ‘open’ to serve corporate interests, the AI industry should align with established open source principles to create a genuinely open ecosystem.”