Open-source AI and the Act

Providers of free and open-source AI are released from some of the Act's obligations, on the reasoning that open models are more transparent by nature and that openness serves research and competition. The relief is aimed at the people releasing the models.

It is a genuine carve-out, but a specific one. It softens certain documentation and transparency duties for open releases; it does not switch the Act off for anything built with them.

The relief falls away in the cases that matter most. It does not apply when a system is high-risk, when it is a prohibited use, when the transparency duties of Article 50 are triggered, or for the most capable general-purpose models judged to carry systemic risk.

In other words, openness helps at the low-risk end and stops mattering at the high-risk end, which is exactly where the heavy duties live.

Treating a model as exempt simply because it is open-source is a common and costly misread. The open-source status of the model under a tool says very little about your obligations.

The use still decides the rules. An open model put to a high-risk job carries the same high-risk duties as a closed one put to the same job.

Your duties come from how you use a tool, not from whether the model beneath it is open or closed. An open model used in a high-risk way is still a high-risk use for you, with the full deployer obligations.

So classify open and closed tools the same way: by what you do with them. The licence on the model is not a shortcut around the Act.