This week, Google released a family of open AI models, the Gemma 3, which quickly received praise for its impressive efficiency. But as a number of Developers Gemma 3’s license makes X lament that commercial use of the model becomes a risky claim.
This is not a problem unique to Gemma 3. Companies like Meta also apply customization of their publicly available models, non-standard licensing terms, and legal challenges to the company. Some companies, especially smaller businesses, are concerned that Google and others can “pull a carpet” by asserting more heavy terms.
“The restrictive and inconsistent licensing of the so-called ‘open AI model’ is creating serious uncertainty, especially for commercial adoption,” said Nick Vidal, community leader of the Open Source Initiative. Long-term operational institutions Designed to define and “steward” all open source things, tell TechCrunch. “Although these models are sold as open, actual terms impose various legal and practical barriers that prevent businesses from integrating them into products or services.”
The reason for open model developers is to publish models under a proprietary license, not industry-standard options Apache and Mit. For example, AI launch commonality Very clear The intention about it supports the work that science (but not business) does on its model.
However, Gemma and Meta’s camel licenses have particular restrictions that limit how companies can use these models without worrying about legal retaliation.
For example, Prohibit developers Improve any model except Camel 3 or “derived works” from using the “output or results” of the Llama 3 model. It also prevents companies with more than 700 million monthly active users from deploying Llama models without first receiving special additional licenses.
Gemma’s license Usually less heavy. But this does grant Google the right to “restrict (remote or otherwise) Gemma’s right”, which Google believes is a violation of the company Prohibited Use Policy or “Applicable laws and regulations.”
These terms apply not only to the original camel and Gemma models. Camel or Gemma-based models must also comply with the Camel and Gemma licenses respectively. In the case of Gemma, this includes a model trained on the synthetic data generated by Gemma.
Florian Brand, research assistant at the German Center for Artificial Intelligence Research, believes that although What kind of technical executives would you believe – Licenses such as Gemma and Llama “can’t reasonably be called ‘open source’.”
“Most companies have a set of approved licenses, such as Apache 2.0, so any custom license is a lot of hassle and money,” Brand told TechCrunch. “Small companies without legal teams or attorney funds will stick to the model of having a standard license.”
Brand notes that AI model developers with custom licenses, such as Google, have not actively enforced their terms. However, he added that the threat is usually enough to prevent adoption.
“These restrictions have had an impact on the AI ecosystem, even on AI researchers like me,” Brand said.
Han-Chung Lee, director of machine learning at Moody’s, agreed that custom licenses such as Gemma and Llama accompanying make the model “unavailable” in many business scenarios. The same goes for Eric Tramel, a scientist at AI startup Gretel.
“Model-specific licenses make model derivatives and distillations specific engravings, which draw attention to kickbacks,” Tramel said. “Imagine a business that specializes in producing models for customers. What license should Llama’s Gemma-Data fine tuning have? What impact will have on all downstream customers?”
Tramel said the most worry about deployment is that the model is a Trojan horse.
“Model casting can be launched [open] Models, waiting for what business cases to develop using these models, and then bringing them into successful verticals through ransomware or legal development. “He said. “For example, Gemma 3 looks like a solid release and can have a wide range of impacts. But due to its licensing structure, the market cannot adopt it. Therefore, businesses may stick to weak and reliable Apache 2.0 models. ”
It should be clear that, despite restrictive permissions, some models have achieved extensive allocation. For example, camels have always been Downloaded hundreds of millions of times And built into products from major companies including Spotify.
Yacine Jernite, head of machine learning and society at AI startup Hugging Face, said they could be more successful if they get permission. Jernite calls on providers like Google to open an open license framework and to work with users “more directly” with users on widely accepted terms.
“Given the lack of consensus on these terms, and the many basic assumptions have not been tested in court, all of this is mainly the declaration of intentions for these actors,” said Genet.[But if certain clauses] Explained too broadly, many excellent jobs will find themselves on an uncertain legal basis, which is especially frightening for organizations that build successful commercial products. ”
Vidal said there is a pressing need for AI model companies to be free to integrate, modify and share without worrying about sudden licensing changes or legal ambiguity.
“The current landscape of AI model licensing is full of confusion, restrictive terms and misleading openness claims,” Vidal said. “Instead of redefining ‘Open’ to fit the interests of the company, the AI industry is aligned with established open source principles to create a truly open ecosystem.”
