Model licenses “ open ” often bear restrictions

This week, Google published a family of open AI models, Gemma 3, which quickly praised their impressive efficiency. But as number of developers Deplored on X, the Gemma 3 license makes the commercial use of models a risky proposition.

This is not a unique problem in Gemma 3. Companies like Meta also apply personalized and non -standard license conditions to their openly available models, and the terms have legal challenges for companies. Some companies, in particular smaller operations, fear that Google and others can “pull the carpet” on their business by affirming the most expensive clauses.

“The restrictive and inconsistent license of the so -called” open “models of IA creates significant uncertainty, in particular for commercial adoption,” said Nick Vidal, head of the Open Source Initiative community, a long -standing institution Aiming to define and “steward” all that is open source, told Techcrunch. “Although these models are marketed as open, the real terms impose various legal and practical obstacles which dissuade companies from integrating them into their products or services.”

Developers of open models have their reasons to publish models under owner licenses as opposed to standard industry options as Apache and with. AI Startup Cohere, for example, was clear About its intention to support scientific work – but not commercial – above its models.

But the LLAMA licenses of Gemma and Meta in particular have restrictions that limit the way companies can use models without fear of legal reprisals.

Meta, for example, Prohibited developers The use of the “output or results” of the LLAMA 3 models to improve any model in addition to LLAMA 3 or “Derived works”. It also prevents companies with more than 700 million monthly active users from deploying LLAMA models without first obtaining a special additional license.

Gemma license is generally less restrictive. But this grants Google the right to “restrict (remotely or otherwise) the use of Gemma which, according to Google, is in violation of the company Prohibited policy of use or “applicable laws and regulations”.

These terms do not only apply to the original models of Lama and Gemma. Llama or Gemma models must also join Llama and Gemma licenses, respectively. In the case of Gemma, this includes models formed on the synthetic data generated by Gemma.

Florian Brand, research assistant at the German research center for artificial intelligence, believes that – despite What the giant leaders of technology would make you believe – Licenses like Gemma and Llama “cannot reasonably be called” open source “.”

“Most companies have a set of approved licenses, such as 2.0 Apache, so any personalized license is a lot of problems and money,” Brand in Techcrunch told. “Small companies without legal teams or money for lawyers will stick to models with standard licenses.”

Brand noted that the developers of AI models with personalized licenses, such as Google, have not yet applied their conditions aggressively. However, the threat is often sufficient to dissuade adoption, he added.

“These restrictions have an impact on the IA ecosystem – even on IA researchers like me,” said Brand.

Han-Chung Lee, director of automatic learning at Moody’s, is suitable that personalized licenses such as those attached to Gemma and Llama make “non-usable” models in many commercial scenarios. Eric Tramel, a staff applied a scientist to the Gretel startup.

“Licenses specific to the model make specific races for model derivatives and distillation, which causes worries concerns,” said Tramel. “Imagine a company that specifically produces setting models for their customers. What license was a Gemma-Data refined from Lama? What would be the impact for all their customers downstream? »»

The scenario that deploys most of fear, said Tramel, is that the models are somehow a Trojan horse.

“A model foundry can turn off [open] Models, wait until you see which profitability analyzes are developing using these models, then spawning a strong path in verticals successful by extortion or law, “he said. “For example, Gemma 3, by all appearances, seems to be a solid release – and which could have a broad impact. But the market cannot adopt it because of its license structure. Thus, companies will probably stick to Apache 2.0 models can be lower and less reliable. »»

To be clear, some models have carried out a general distribution despite their restrictive licenses. Lama, for example, was downloaded hundreds of millions of times and integrated into products from large companies, including Spotify.

But they could be even more successful if they were authorized with a license, according to Yacine Jernite, responsible for automatic learning and the company at the AI ​​startup hugging the face. Jernite called suppliers like Google to switch to license frames or to open license frames and to “collaborate more directly” with users in widely accepted terms.

“Given the lack of consensus on these terms and the fact that many of the underlying hypotheses have not yet been tested before the courts, everything mainly serves as a declaration of intention of these actors,” said Jernite. “”[But if certain clauses] are interpreted too largely, many good work will be on an uncertain legal ground, which is particularly frightening for organizations that build successful commercial products. »»

Vidal said there was an urgent need for AI models, companies can freely integrate, modify and share without fear of sudden license changes or legal ambiguity.

“The current AI model license landscape is riddled with confusion, restrictive terms and misleading opening claims,” ​​said Vidal. “Instead of redefining” open “to adapt to the interests of companies, the AI ​​industry should align with the open source principles established to create a really open ecosystem.”