Checkpoint: Can This Release Claim the Exception?

The question now is not whether a project calls itself open-source. It is whether the licence permissions, public model information, and Article 53 boundaries line up in one release. This checkpoint connects the two-part eligibility test with the baseline provider duties already established, so a reader can make a provisional judgment from concrete facts, identify what remains, and stop where the evidence runs out.

The scenarios below are deliberately similar. Each changes one important fact. That makes the legal mechanism visible: the exception is conditional, its scope is limited to specified duties, and systemic risk can override it even when the openness conditions appear satisfied.

Scenario one: the conditions appear to align

Imagine a provider releases a GPAI model under terms that allow people to access it, use it, modify it, and distribute it. The provider also makes the parameters, including the weights, publicly available and publishes information about the model’s architecture and usage. Assume further that the available facts do not establish that the model has systemic risks.

On those facts, the release appears to satisfy the two conditions in Article 53(2). The licence supplies the four named permissions. The public release package supplies the parameters and the architecture and usage information named by the rule. The careful conclusion is that the provider may be able to rely on the exception for the specified documentation duties.

That conclusion is intentionally narrower than “the model is exempt.” The words “may be able” acknowledge both the factual nature of the test and the rule’s scope. The release has crossed the eligibility threshold for a particular adjustment, not received a general certificate of freedom from the AI Act.

The distinction matters because the provider’s release still has a legal context. The model is made available by an actor who may be a GPAI model provider even though the model is free. The release remains connected to the provider duties described in Article 53, and the model may later be modified or integrated by other actors. Qualification at the first stage is a starting point for further analysis, not the end of it.

Scenario two: one missing permission changes the result

Now change only the licence. Suppose the provider makes the weights public and permits recipients to access and use them, but the licence does not allow recipients to modify the model. The release may be technically accessible and useful, but the Article 53(2) condition is not established because modification is one of the permissions named by the rule.

Change the fact again. Suppose modification is permitted, but redistribution of the model or a modified version is restricted. The same problem appears in a different form. A recipient may be able to experiment locally, yet the release would not clearly provide the full set of access, usage, modification, and distribution permissions required by the text.

These examples do not establish how every licence with conditions should be treated.