GPAI deadline page for providers of general-purpose AI models, including foundation models and LLMs offered to others.
What the obligation actually requires
GPAI duties are aimed at providers of general-purpose AI models: the companies that make a model available for other developers, integrators, or business users to build on. In practical terms, the work is about knowing what model you are providing, keeping technical documentation in a form that can be shared when needed, giving downstream providers enough information to use the model responsibly, and maintaining a copyright policy. If the model has systemic-risk exposure, the work does not stop at a model card. It needs a deeper internal process for documenting risks, changes, and how the model is released.
Who usually triggers it
A company that trains and offers a foundation model through an API is the clearest case. An open-weight model publisher can also be in the conversation if it is providing a general-purpose model for others to adapt. A vertical software company may trigger this if it is not merely using someone else’s model, but packaging and offering its own general-purpose model layer to customers or partners.
The common misunderstanding
The common mistake is assuming this only applies to consumer chatbots. The deadline data is about model providers, not only chatbot operators. A business can avoid a chatbot UI entirely and still have GPAI exposure if it provides a model that downstream teams can use across many tasks. The opposite mistake also happens: a normal business user of a third-party model may have other AI Act duties, but it is not automatically the GPAI model provider.
What to do next
Start by separating three roles: model provider, downstream provider, and deployer. Then list every model your company trains, fine-tunes, hosts, or exposes to others. For each one, identify what documentation exists, what information downstream users receive, and who owns copyright-policy review. If you may have systemic-risk exposure, escalate that question early. Do not wait for procurement or product launch to discover that the model documentation is incomplete.