Meta released Llama 4 last week to the usual fanfare about open source AI advancing the state of the art. Except Llama 4 isn’t open source in any meaningful sense. The weights are available for download, sure, but the training data is proprietary, the training methodology is undisclosed, and the license prohibits commercial use above certain thresholds without Meta’s permission.

This is corporate-controlled AI with a marketing strategy that appropriates open source language. It’s effective marketing—developers love the idea of open models—but it’s fundamentally different from actual open source software.

What Open Source Actually Means

The Open Source Initiative has clear criteria: free redistribution, source code availability, derived works permitted, no discrimination against persons or fields of endeavor. Traditional open source software like Linux, PostgreSQL, or Python meets these standards completely.

AI models claiming to be “open source” typically fail on multiple criteria. You can’t audit the training process, can’t replicate the model from scratch, can’t verify that the training data was ethically sourced, and often face commercial licensing restrictions.

Llama, Mistral, Stable Diffusion—all positioned as open alternatives to proprietary models, all controlled by entities with commercial interests that limit genuine openness. The weights are available, but weights without transparency on training methodology are just free-to-use proprietary technology.

Why Corporations Like “Open Source” AI

Releasing model weights publicly achieves several corporate objectives that have nothing to do with openness:

Developer ecosystem capture: Get developers building applications on your model architecture, creating lock-in and dependency Regulatory positioning: Look collaborative and pro-innovation compared to closed competitors, influence regulatory frameworks Training data acquisition: Open models attract users who generate data that improves future proprietary versions Compute cost externalization: Let the community handle inference costs and edge case discovery

Meta doesn’t release Llama because they believe in open source principles. They release it because a vibrant Llama ecosystem benefits Meta strategically. The moment that calculation changes, the “openness” disappears.

The Truly Open Alternative

There are genuinely open AI projects, but they’re smaller and less promoted. EleutherAI releases models with fully documented training processes and transparent data sources. The models aren’t as performant as corporate releases, partly because genuine openness comes with constraints that limit scale.

To train a truly open model, you need open datasets (which excludes most web scraping), disclosed compute resources (which reveals costs competitors can undercut), and transparent methodology (which means no proprietary optimizations). That’s a significant competitive disadvantage.

Corporate “open source” AI avoids those constraints. They don’t disclose what data they trained on, so you can’t audit for copyright violations or bias sources. They don’t explain their training process, so you can’t identify where performance comes from or replicate results independently.

Regulatory Implications

This matters beyond philosophical debates about terminology. Policymakers are making decisions about AI regulation based on false distinctions between “open” and “proprietary” models. If open models are treated as lower-risk or exempt from certain requirements, corporate AI labs will ensure their products qualify as “open” regardless of actual transparency.

The EU AI Act includes provisions that differentiate open source AI systems. If that distinction relies on simply releasing model weights rather than genuine openness criteria, it creates a loophole large enough to drive Meta’s entire AI strategy through.

Australia’s looking at similar regulatory frameworks. We need to define “open source AI” rigorously, with requirements that go beyond weight availability to include training transparency, data disclosure, and genuine reproducibility.

What Users Should Demand

If you’re building applications on AI models marketed as “open source,” understand what you’re actually getting:

• Weights: Yes, available for download
• Architecture: Usually documented
• Training data: Almost never disclosed
• Training process: Proprietary and undisclosed
• Reproducibility: Impossible without massive undisclosed resources
• License restrictions: Often present, limiting commercial use

That’s a useful resource—free-to-use model weights have value—but it’s not open source in the way that term has meant for 25 years of software development.

The Path Forward

Either we enforce the term “open source AI” to mean genuinely open (disclosed training data, reproducible methodology, unrestricted licensing), or we abandon the term and call these models what they are: freely available proprietary technology.

I’d prefer the former. Open source software transformed technology by enabling genuine transparency, collaboration, and distributed innovation. AI needs that same foundation, not a corporate-friendly simulacrum that appropriates the language while abandoning the principles.

Meta, Mistral, and Stability AI aren’t going to police themselves. It’s on the development community and policymakers to demand actual openness if we’re going to use that term. Otherwise, we’re letting corporations define “open source” to mean whatever serves their strategic interests.