This week, Google released a family of open AI models, Gemma 3. However, as many developers lament the X, the Gemma 3 license makes the commercial use of the model a dangerous proposal.
It’s not a problem specific to Gemma 3. Companies like Meta also apply custom and non-standard licensing terms to openly available models, and the terms present legal challenges to the companies. Some companies, especially small businesses, are worried that by asserting more troublesome clauses, companies like Google can “pull a rug” into their businesses.
“The restrictive and inconsistent licensing of the so-called “open” AI model creates significant uncertainty, especially for commercial adoption,” Nick Vidal, community head of open source initiatives, told TechCrunch that it is a long-term institution aimed at defining and “stewarding” everything in open source. “Though these models are sold as open, the actual terminology places a variety of legal and practical hurdles that prevent companies from integrating into products and services.”
Open model developers have a reason to release models under their own licenses, as opposed to industry standard options such as Apache and MIT. For example, AI startup Cohere has made clear about its intention to support it on a model, not a scientific (not commercial, not commercial, and not commercial).
However, the llama licenses in particular for Gemma and Meta have restrictions that limit the way companies can use the model without fear of legal retaliation.
For example, Meta prohibits developers from using “output or result” from Llama 3 models to improve models other than Llama 3 or “derived work.” It also prevents over 700 million active users from deploying the Llama model every month without first obtaining a special additional license.
Gemma licenses are generally less burdensome. However, Google grants Google the right to “restrict remote or other use” of Gemma, and we believe that it is in violation of the company’s prohibited use policy or “appropriate laws and regulations.”
These terms do not apply only to the original Llama and Gemma models. Models based on Llama or Gemma must comply with the Llama and Gemma licenses, respectively. For Gemma, it includes models trained with synthetic data generated by Gemma.
Florian Brand, a research assistant at the German Center for Artificial Intelligence Research, believes that despite the huge high-tech executives believe you, he believes in licenses such as Gemma and Llama “can’t reasonably say “open source””.
“Custom licenses are a lot of trouble and money because most companies have a set of approved licenses, such as Apache 2.0,” the brand told TechCrunch. “Small businesses with no money for legal teams and lawyers stick to models with standard licenses.”
The brand noted that custom licensed AI model developers like Google have not yet actively enforced terms. However, he added that threats are often sufficient to block adoption.
“These restrictions impact the AI ecosystem, even for AI researchers like me,” Brand said.
Hancheon Lee, director of machine learning at Moody’s, agrees that custom licenses like those attached to Gemma and Lama will prevent the model from being “usable” in many commercial scenarios. So does Eric Tramel, a staff applied scientist at AI Startup Gretel.
“Model-specific licenses make certain carve-outs for model derivatives and distillations. This raises concerns about clawbacks,” Tramel says. “Imagine a business that is specifically producing fine-tuning models for its customers. What license should tweak Rama’s Jemmadata tweaks have? What impact will it have on all customers downstream?”
The scenario that deployers fear most is that models are some kind of Trojan horse.
“Model Foundries can output (open) models, wait for which business cases to occur using those models, then register them strongly vertically, either by fear or by law,” he said. “For example, Gemma 3 with all appearances looks like a solid release. And it could have a wide range of impact. However, the market cannot adopt it because of the licensing structure. So companies will probably stick to the weak and unreliable Apache 2.0 model.”
To be clear, certain models achieved wide distribution despite restrictive licenses. For example, Llama has been downloaded hundreds of millions of times and is incorporated into its products from large companies, including Spotify.
But according to Yacine Jernite, head of machine learning and society for AI startup hug faces, they could be even more successful if they were allowed. Jernite has called on providers like Google to move to open the licensing framework and move to “more direct cooperation” with users on widely accepted terms.
“Given the lack of consensus on these terms and the fact that many of the underlying assumptions have not yet been tested in court, it all serves primarily as a declaration of intent from those actors,” Jernite said. “If (but certain clauses) are interpreted too widely, many good jobs will find themselves on an uncertain legal foundation. This is especially scary for organizations building successful commercial products.”
Vidal said there is an urgent need for AI model companies that can freely integrate, modify and share without fear of sudden license changes or legal ambiguity.
“The current landscape of AI model licensing is full of misleading claims of confusion, restrictive conditions and openness,” Vidal said. “Instead of redefineing ‘open’ to fit the profits of businesses, the AI industry must create a truly open ecosystem in line with established open source principles. ”