In the rapidly evolving world of cryptocurrency and blockchain, the concept of ‘open’ and ‘decentralized’ is paramount. But what happens when the ‘open’ label is applied to another groundbreaking technology like Artificial Intelligence? This week’s buzz around Google’s Gemma 3, touted as a family of open AI models , highlights a concerning trend: not all that is labeled ‘open’ truly empowers users, especially in the commercial sphere. Just as you scrutinize the fine print in crypto contracts, the licenses governing these AI models demand a closer look. The Illusion of Openness: Are AI Model Licenses Truly Open? Google’s Gemma 3, celebrated for its efficiency, quickly faced scrutiny over its licensing terms. Developers flocked to X (formerly Twitter) to voice their concerns, pointing out that the license makes commercial AI use a risky endeavor. This isn’t an isolated incident. Meta, with its Llama models, also employs custom, non-standard licenses, creating a legal maze for businesses. Smaller companies, in particular, fear a potential ‘rug pull’ – a sudden enforcement of onerous clauses by tech giants like Google and Meta. Nick Vidal, head of community at the Open Source Initiative, a respected authority on open source, aptly summarizes the situation: “The restrictive licenses and inconsistent licensing of so-called ‘open’ AI models are creating significant uncertainty, particularly for commercial adoption. While these models are marketed as open, the actual terms impose various legal and practical hurdles that deter businesses from integrating them into their products or services.” Why Aren’t Open AI Models Truly Open Source? Unlike standard open-source licenses like Apache and MIT, AI model developers often opt for proprietary licenses for specific reasons. Cohere, an AI startup, openly states its intention to support scientific research but limit commercial AI use . However, the licenses from Gemma and Meta’s Llama models go further, imposing restrictions that can deter businesses: Meta’s Llama 3 License: Prohibits using model outputs to improve any model other than Llama 3 or its derivatives. Requires a special license for companies with over 700 million monthly active users. Google’s Gemma License: Grants Google the power to remotely restrict usage if they deem it violates their policies or applicable laws. These restrictions extend beyond the original models. Derivatives of Llama and Gemma, even those trained on synthetically generated data, must adhere to the same licensing terms. The Chilling Effect of Custom Licenses on the AI Ecosystem Florian Brand, a research assistant at the German Research Center for Artificial Intelligence, argues that licenses like Gemma and Llama’s cannot be genuinely considered ‘open source.’ He highlights the practical difficulties for businesses: Legal and Financial Burden: Custom licenses deviate from standard, approved licenses like Apache 2.0. This creates “trouble and money” for companies, especially smaller ones without dedicated legal teams. Deterrent to Adoption: Even without aggressive enforcement, the mere threat of complex legal terms discourages adoption, pushing businesses towards models with simpler, standard licenses. Brand emphasizes, “These restrictions have an impact on the AI ecosystem — even on AI researchers like me.” Commercial Unviability and the Trojan Horse Fear Han-Chung Lee from Moody’s and Eric Tramel from AI startup Gretel echo Brand’s concerns, stating that custom licenses render models like Gemma and Llama “not usable” in many commercial scenarios. Tramel points out the ambiguity around derivative models, questioning the licensing implications for fine-tuned models and their downstream customers. The ultimate fear, as Tramel describes it, is that these models are a “trojan horse.” Model foundries could release seemingly open AI models , observe successful commercial applications, and then leverage restrictive licenses for extortion or legal battles. He uses Gemma 3 as an example: despite its potential, its restrictive license structure hinders market adoption, potentially pushing businesses towards less capable but permissively licensed alternatives. The Untapped Potential of Truly Open Models Despite licensing hurdles, models like Llama have seen widespread adoption, even by major corporations like Spotify. However, Yacine Jernite from Hugging Face argues that their success could be even greater with truly permissive licenses. He urges providers like Google to embrace open license frameworks and collaborate with users on broadly accepted terms. Jernite warns that the current lack of legal clarity and consensus around these custom licenses creates uncertainty, especially for businesses building commercial products. Vidal underscores the urgent need for open AI models that companies can freely use, modify, and share without legal ambiguity or the fear of sudden license changes. Towards a Genuinely Open AI Ecosystem The current state of AI model licenses is, as Vidal puts it, “riddled with confusion, restrictive terms, and misleading claims of openness.” Instead of redefining ‘open’ to suit corporate interests, the AI industry should genuinely embrace established open-source principles to foster a truly open and innovative AI ecosystem . Just as the crypto world champions decentralization and open access, the AI space needs to move beyond the illusion of openness and towards genuine empowerment for developers and businesses. To learn more about the latest AI market trends, explore our articles on key developments shaping AI features and institutional adoption.
