Meta's AI chief says open source licensing is broken for LLMs

When Meta launched Llama 2 in July, it marketed the model as a massive win for open source, putting frontier-level AI into the hands of developers for free. But old-school open-source purists have spent the last few months waving red flags, arguing that Meta is effectively co-opting the term.

In an interview published October 30 by The Verge, Meta’s head of AI research and FAIR leader Joelle Pineau pushed back. Her core argument is straightforward: Meta’s restrictions are necessary, and traditional software licenses simply were not built to handle the complexities of modern generative AI.

The fine print behind Meta's generosity

According to the Open Source Initiative (OSI), open source requires unrestricted redistribution, full access to source code, and zero discrimination against specific fields of use or commercial scale.

Meta’s Llama 2 license breaks those rules in two big ways. First, any product with over 700 million daily active users must request a custom commercial license and pay Meta. Second, developers are explicitly forbidden from using Llama 2’s outputs to train competing models.

Researchers at Radboud University publicly called Meta’s branding misleading in IEEE Spectrum, and developer forums echoed the complaint. Pineau acknowledged these limits, but defended them as an essential compromise to protect Meta’s business from rival tech giants.

She pointed out that Meta has a proven track record with true open source, citing PyTorch, which Meta released openly in 2016 without these guardrails. But large language models, she argued, introduce unprecedented risks around training data and copyright liabilities that old software licenses never accounted for.

OSI executive director Stefano Maffulli acknowledged the dilemma, telling The Verge that the organization is actively working on a formal definition of open-source AI to address these gaps without abandoning community principles.

Why it matters

We think this feud exposes an unavoidable clash between corporate self-preservation and traditional hacker ethics.

For everyday developers and startups, Llama 2 is still an incredible gift. Unless you are building an app that rivals Instagram in scale, you can deploy it commercially without paying a dime. That makes it far more accessible than closed APIs from OpenAI or Google. Debating licensing taxonomy might feel like academic hair-splitting when you just want to build a product.

For the broader industry, however, the stakes are massive. As closed-source leaders like OpenAI double down on secrecy, Meta wants the prestige and talent pipeline of an open champion while secretly rewriting the rulebook.

Meta's compromise looks commercially clever, but it sets a dangerous precedent. If tech giants are allowed to slap the open-source label on models packed with proprietary restrictions and commercial moats, the term loses its meaning, leaving the open-source community divided just as AI standards are being permanently etched into law.