The empty top of open

On 29 May, meeting in Paris under the French G7 presidency, the seven governments' digital ministers signed off on a 'Vision on AI openness' — the first time governments, rather than NGOs or standards bodies, have put a shared vocabulary behind the word 'open.' It sorts models onto four rungs by what a lab actually releases: at the bottom, 'Weights Available AI' (you can inspect the weights, but the licence isn't an open-source one); then 'Open Weights AI' (weights and code under a genuine open licence); then 'Open Source AI' (adds the training code); and at the top, 'Open Source AI with Open Data' (everything, training data included).

The document is a non-binding vision — aspirational language, not a rule — and the Open Source Initiative helped draft it, importing distinctions its own definition already drew. So the story isn't the ladder; it's who stands on it. When one analyst applied the G7's own rungs to about forty competitive models, they split between the bottom two — and not a single one reached either top tier. The two most-open categories the ministers codified are, at the moment of writing, empty.

That inverts the press-release story. Rather than a standard that shames 'open' models into opening further, the taxonomy reads as governments quietly settling on 'open weights' as the working definition of open AI — the way 'open source' once displaced 'free software' — while parking a purer 'open source' label that market reality has already vacated. A model like Llama, marketed as open, lands on the bottom rung; the vocabulary now exists for a regulator or the EU AI Act to say so precisely, even if today it says nothing binding at all.