‘Open’ model licenses often carry concerning restrictions

This week, Google has released a family of Open AI models, Gemma 3, who quickly led praises for their impressive efficiency. But like one number by developers Complainted about X, the license from Gemma 3 makes commercial use of the models a risky proposition.

It is no problem that is unique for GEMMA 3. Companies such as Meta also fit customized, non-standard license conditions on their openly available models, and the conditions present legal challenges for companies. Some companies, especially smaller activities, are concerned that Google and others “can pull the rug” to their company by asserting the more heavy clauses.

“The limiting and inconsistent licenses of so-called ‘open’ AI models creates considerable uncertainty, in particular for commercial acceptance,” Nick Vidal, head of the community in the Open Source Initiative, A long -term institution Strive to define all things open source and “rose”, WAN said. “Although these models are brought open to the market, the actual conditions lay different legal and practical obstacles on those companies prevent them from integrating into their products or services.”

Open Model Developers have their reasons for releasing models under their own licenses in contrast to industrial standards such as such as such as Apache and MIT. AI Startup Cochere for example, has been clear About his intention to support scientific – but not commercial – work on top of his models.

But in particular the LLA licenses from Gemma and Meta have limitations that limit the ways in which companies can use the models without fear of legal reprisals.

Meta for example, Prohibits developers From the use of the “Output of Results” of Llama 3 models to improve each model in addition to Llama 3 or “Derivative Works”. It also prevents companies with more than 700 million monthly active users to implement LLAMA models without first obtaining a special, extra license.

Gemma’s license Is generally less difficult. But it does give Google the right to “limit” use (remote or otherwise) use of Gemma, according to Google is contrary to the company Forbidden user policy or “applicable laws and regulations.”

These conditions do not only apply to the original Lama and Gemma models. Models based on Llama or Gemma must also adhere to the Lama and Gemma Licencies respectively. In the case of Gemma, that models comprises trained on synthetic data generated by Gemma.

Florian Brand, a research assistant at the German research center for artificial intelligence, believes that – despite What technical gigantic execs would you like to believe – Licenses such as Gemma and Lama’s’ cannot reasonably be called ‘open source’.

“Most companies have a number of approved licenses, such as Apache 2.0, so every adjusted license is a lot of problems and money,” Brand told WAN. “Small companies without legal teams or money for lawyers will adhere to models with standard licenses.”

Brand noted that AI model developers with adapted licenses, such as Google, have not yet forced their conditions aggressively. However, the threat is often sufficient to deter the adoption, he added.

“These limitations influence the AI ​​ecosystem -even on AI researchers like me,” said Brand.

Han-Chung Lee, director of Machine Learning at Moody’s, agrees that adapted licenses such as those associated with Gemma and Lama make the models “not usable” in many commercial scenarios. This also applies to Eric Tramel, a staff used at AI Startup Gretel.

“Model-specific licenses make specific carve-outs for model derivatives and distillation, which causes concern about Clawbacks,” said Tramel. “Imagine a company that produces specific model deviations for their customers. What license should a GEMMA data finish from Llama have? What would be the impact for all their downstream customers? “

The scenario that the implementation fears the most, Tramel said, is that the models are a kind of Trojan horse.

“A model foundry can release [open] Models, wait to see which business cases are developing with the help of those models, and then their way to successful verticals through extortion or lawfare, “he said. “For example, Gemma 3 seems to be a solid release in all performances – and one that can have a broad impact. But the market cannot take over because of the license structure. So companies are likely to stay with perhaps weaker and less reliable Apache 2.0 models. “

To be clear, certain models have reached a widespread distribution despite their limiting licenses. For example, Lama has been downloaded hundreds of millions of times and built into products from large companies, including Spotify.

But they can be even more successful if they had a permanent license, according to Yacine Jernite, head of Machine Learning and the society at AI Startup Hugging Face. Jernite called on providers such as Google to open license frames and “to work more directly together” with users on widely accepted conditions.

“Given the lack of consensus on these conditions and the fact that many of the underlying assumptions have not yet been tested in the courts, it all serves in the first place as a statement of those actors,” said Jernite. ‘[But if certain clauses] Being interpreted too broadly, a lot of good work will be on uncertain legal land, which is particularly scary for organizations that build successful commercial products. “

Vidal said there is an urgent need for AI models companies that can freely integrate, change and share without fearing with sudden licensing changes or legal ambiguity.

“The current landscape of AI Model Licensing is full of confusion, restrictive conditions and misleading claims of openness,” said Vidal. “Instead of redefining ‘open’ for company interests, the AI ​​industry must be in accordance with established open source principles to create a real open ecosystem.”