'Open' in AI doesn't mean what you think: how to check it yourself

Picture a developer in Madrid who has spent weeks building a prototype on top of Llama 4, Meta's model. Every headline calls it an "open-weight model": you can download it, run it on your own server, modify it as you like. One day you open the full license — not the summary circulating on social media, the actual document — and find this sentence, applied to the model's multimodal versions: usage rights "are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union." That is not a technical block, and it is not a side effect of the EU's AI Act. It is a clause Meta wrote into its own contract. "Open," for you, had a border nobody had pointed out.

This happens to more people than you'd think, and it isn't unique to Meta. "Open" has become AI's favorite adjective — every lab uses it, and almost none of them mean the same thing by it — and mixing up its variants has real consequences: projects that assume total freedom and run into a scale clause, teams that discover too late that "open weights" includes neither the data nor the recipe the model was trained with. The skill you need isn't memorizing which model is "the most open" this month — that changes constantly — it's knowing how to read any openness claim and check it yourself, with the document in front of you.

What "open" actually means in AI

For decades, "open source" had a stable definition: the one from the Open Source Initiative (OSI), built for conventional software. Artificial intelligence put that definition under strain. What does it mean to "open" a system when what matters most isn't the code but the weights — the billions of numbers the model learned during training — and the data those numbers came from? The OSI answered in October 2024 with the Open Source AI Definition (OSAID) 1.0, which requires three components to be released under terms that let anyone use, study, modify and share the system without asking permission:

• The model's parameters (the weights, or an equivalent configuration).
• The complete code used to train and run it.
• Sufficiently detailed information about the training data, so that a skilled person could build a substantially equivalent system.

Notice how demanding that third point is: it isn't enough to say "trained on web data"; you need enough detail to rebuild it. It's the piece almost no major lab meets — and the easiest one to miss when a headline simply says "open model."

The Llama case: weights, yes. Recipe, no.

Llama is the most visible example, so it works as a map for the rest. Its license grants broad rights to use, modify and redistribute the model — but with conditions no genuine open-source license carries. The most-cited one: if your company's products exceed 700 million monthly active users, you cannot use Llama without requesting a separate license from Meta, which the company "may grant to you in its sole discretion." And as you saw above, the multimodal versions flatly exclude any individual domiciled in, or company headquartered in, the European Union, regardless of size.

Here a pause on how this piece was researched is worth taking, because it's part of the lesson itself. Several sources circulating online claim Llama's license bans using its outputs to train rival models. Read the license word for word — not a summary, the legal text — and that isn't what it says: if you use Llama's outputs to train another model and distribute it, you have to put "Llama" at the start of its name. That's an attribution requirement, not a ban. The difference matters, and it's exactly the kind of nuance that gets lost when you repeat what someone said the license says instead of opening the document yourself.

What Llama does not include, on the other hand, is the training corpus or the code used to produce it: you get the finished model, not the recipe or the ingredient list. That's why Amanda Brock, CEO of the UK's OpenUK, has been blunt: Llama "doesn't meet the OSI's open source standard", among other reasons, because of the commercial restrictions in its own license.

The counterexample: what genuine openness looks like

To keep the contrast concrete rather than abstract, consider a model that does meet all three OSAID requirements. OLMo 2, from the Allen Institute for AI (AI2), ships with the weights, the training code, intermediate checkpoints, and — unusually — the dataset itself: Dolma, an open corpus of roughly three trillion tokens of web pages, academic papers, code and books, released under a license that allows free reuse. All of it under Apache 2.0, with no user thresholds and no geographic carve-outs. AI2 sums it up in a line that could serve as this article's headline: "fully open science requires more than just open weights."

Nobody is telling you to use OLMo instead of Llama — they're different-capability models for different needs. The point of knowing OLMo exists is different: it gives you the far end of the spectrum to measure any future "openness" announcement against. And the same exercise applies to Mistral, Gemma, DeepSeek, or whichever lab uses the word next — none of the questions below depend on who happens to be leading this month.

Why openness is almost always partial

It helps to understand the incentive behind the pattern. A model's weights are its most expensive asset to produce and its main competitive edge; releasing them at all is already a real concession, and almost always a strategic one — opening the model attracts developers, and the customization or support built around it is where the lab expects to make money. Releasing the full training data is a different kind of step: it exposes editorial choices about what went in and what didn't, copyright exposure, and the possibility that a competitor rebuilds the whole system without paying for the training run. That's why most labs stop at the weights. It isn't necessarily bad faith; the last mile of openness costs far more than the first two.

The three questions worth asking

This is the part that stays useful a year from now, when it's a different lab making the "open" claim. Before repeating the adjective, check:

1. What exactly is being released? Weights, training code, data — or just the first one? Every missing piece is something you cannot audit or reproduce on your own.
2. Under what license, and what does it condition? Look for user thresholds, country or sector exclusions, and mandatory attribution clauses. They're almost never in the announcement; they're in a separate document you have to open.
3. Does it meet a formal definition, or is it a marketing adjective? The OSI's Open Source AI Definition is the public yardstick available today; if a model doesn't meet its three components, "open" describes an intention, not a verifiable fact.

None of these three questions require knowing how to code. They require reading a legal document instead of a headline, and anyone can do that with ten minutes and the right link.

Check it yourself

You don't need to take this article's word for it: the Llama 4 license is published in full, so is the Open Source AI Definition, and the OLMo 2 announcement explains exactly what's released and under what terms — all three linked above, right where they're cited. Next time a headline says an AI model "is open," that's the ten-minute check that separates actually knowing from having read it somewhere.