This week, Google released a family of open AI models, Gemma 3, which quickly gained praise for their wonderful excellence. But as a number of Developing Mourning X, the Gemma 3 license produces commercial use of models of a risky proposal.
This is not a problem that is unique to Gemma 3. Companies like meta are also applying custom, non-standard licensing terms to their open available models, and terms present legal challenges for companies. Some companies, especially smaller operations, remember that Google and others can “pull the rug” in their business by asserting a heavier clause.
“The restriction and uneven licensing of the so-called 'open' AI models creates significant uncertainty, especially for commercial adoption,” Nick Vidal, community leader in the open source initiative, a Long-running institution It aims to define and “steward” all the open resources, told Techcrunch. “While these models are sold as tomorrow, the actual terms impose a variety of legal and practical barriers that prevent businesses from incorporating them into their products or services.”
Open model developers have their reasons for releasing models under ownership licenses compared to industrial standards such as Apache and included. Ai startup cohere, for example, It became clear Regarding its intent to support the scientific – but not commercial – work on top of its models.
But Gemma and Meta licenses are particularly with restrictions that limit the ways that companies can use models without fear of legal payment.
For example, Meta, The developers are prohibited From the use of “output or results” of Llama 3 models to improve any model besides Llama 3 or “derivative works.” It also prevents companies with over 700 million monthly active users from deploying Llama models without first getting a special, additional license.
Gemma license It is generally not so heavy. But it gives Google the right to “restrict (remote or otherwise) Gemma's use” that Google believes to violate the company Prohibited use policy or “applicable laws and regulations.”
These words do not only apply to the original Llama and Gemma models. Models based on Llama or Gemma must also comply with Llama and Gemma licenses, respectively. In Gemma's case, it includes models trained in synthetic data generated by Gemma.
The Florian Brand, a research assistant at the German Research Center for artificial intelligence, believes that – despite What tech giant execs do you want to believe – Licenses like Gemma and Llama do not reasonably be called 'open resources.' “
“Most companies have a set of approved licenses, such as Apache 2.0, so any custom license has a lot of problems and money,” Brand told TechCrunch. “Small companies without a legal team or money for lawyers will remain in models with standard licenses.”
The brand noted that AI model developers with custom licenses, such as Google, have yet to aggressively implement their terms. However, the threat is often sufficient to prevent adoption, he added.
“These restrictions have an impact on the AI ecosystem – even on AI researchers like me,” Brand said.
Han-Chung Lee, director of Machine Learning Machine Learning, agrees that custom licenses such as Gemma and Llama are attached to “unavailable” models in many commercial scenarios. So did Eric Tramel, a staff member who applied to AI startup Gretel.
“Model-specific licenses make specific carve-outs for model derivatives and distillation, causing concern about clawbacks,” Tramel said. “Imagine a business that specifically makes model fine-tuns for their customers. What license should a gemma-latin fine-tune of Llama? What would be the effect for all of their upstream customers?”
The scenario feared by deployers, Tramel said that the models are a Trojan horse.
“A foundry model can be released [open] Models, wait to see what business cases form with those models, and then strongly bark their way with successful verticals through either extortion or law, ”he said. “For example, Gemma 3, in all displays, seems to be a solid release – and one that can have a broad impact. But the market cannot use it because of its license structure. Thus, businesses are likely to stick to perhaps weaker and less reliable Apache 2.0 models.”
To be clear, some models have achieved extensive distribution despite their restrictions licenses. For example, Llama Downloaded the road — a million times and built on products from major corporations, including Spotify.
But they can be more successful if they are allowed to be licensed, according to Yacine Jernite, head of machine and society study at AI startup embrace. Jernite calls on providers such as Google to move to open license frameworks and “cooperate more directly” with users in widely accepted terms.
“Due to the lack of consensus on these words and the fact that many of the underlying assumptions have not been tested in courts, it all serves as an expression of intent from those actors,” Jernite said. “[But if certain clauses] Extensively interpreted, many good activities will find itself in uncertain legal soil, which is particularly scary for organizations that build successful commercial products. “
Vidal said there is an urgent need for models of AI models to freely include, change, and share without fear of changes in sudden licenses or legal ambiguity.
“The current view of the AI model licensing has survived the confusion, restriction terms, and misleading claims of openness,” Vidal said. “Instead of re -defining 'tomorrow' to adapt to corporate interests, the AI industry must be aligned with the established principles of open resources to create a truly open ecosystem.”
