The European Parliament is building and testing an internal generative AI platform, the EPGenAI Hub, intended to give members and staff sanctioned access to commercial language models rather than leaving them to paste parliamentary material into public chatbots. The platform has been demonstrated to MEPs and staff at workshops and is in beta, and the Parliament's press office has confirmed that the selection of models it offers will be reviewed against European sovereignty considerations. Politico, which first reported the platform, said it could be rolled out to members and staff as early as September, citing sources; that timing has not been formally confirmed by the institution.

What the platform actually does

The hub functions as a controlled gateway in front of several models rather than a single assistant, letting a user choose according to how sensitive the task is. One tier runs on the Parliament's own infrastructure and consists of open-weight models, reported by Politico and subsequently corroborated by multiple outlets to include Meta's Llama, OpenAI's GPT-OSS and Mistral Small. A second tier reaches more capable externally hosted services, reported to include OpenAI's ChatGPT and Anthropic's Claude Sonnet. The design principle is containment: sensitive legislative drafting stays on internally hosted models, while higher-capability external models are available for work where the data classification allows it.

The driver is scale of unmanaged use. Around a fifth of the Parliament's staff, roughly 2,100 people, are reported to use AI tools daily, and committee chairs have raised concerns that members and assistants are leaning on those tools for amendments and other legislative work. The signs cited include a surge in amendments requiring translation, recognisable drafting patterns, and references to laws that do not exist. Brando Benifei, the Socialists and Democrats MEP who led the Parliament's work on the AI Act, said publicly that Parliament has not been immune to AI-generated filler and misleading content, and that MEPs should set an example as the Act's transparency obligations take effect.

The sovereignty problem the hub does not solve

Two limitations are worth stating plainly. The first is compositional: among the models currently reported, only Mistral is developed by a European vendor, although Parliament has indicated the model selection remains under review, which means an institution that has spent two years regulating US technology firms is preparing to run its own legislative workflow substantially on US models. The Parliament's press office has said the model selection will be re-examined on sovereignty grounds, which is an acknowledgement that the current lineup is a starting point rather than a settled answer. The second is that a gateway solves a data-governance problem, not an accuracy one. Routing a prompt through managed infrastructure controls where parliamentary text goes and creates a logging and classification layer; it does nothing to stop a model inventing a legal citation. The hallucinated references that prompted the project are likely to persist unless verification practices evolve alongside the tooling.

What it means for the region

The pattern is directly relevant to Southeast Asian governments and regulators working through the same sequence. Staff adopt commercial AI faster than policy can govern it; the institution discovers official material is flowing into systems it does not control; and the response is a managed internal gateway rather than a ban. For agencies handling material subject to the PDPA or its regional equivalents, the EPGenAI Hub is a useful reference for the architecture — tiered model access keyed to data classification, local hosting for sensitive work, logging at the gateway — and an equally useful reminder that the sovereignty question does not disappear when you build the gateway. It just moves to which models sit behind it, a trade-off that on our reading ASEAN institutions may face even more acutely than Brussels does, since the region currently has fewer domestically developed frontier models to choose from.

Key Takeaways

  • The European Parliament is testing the EPGenAI Hub, an internal platform giving MEPs and staff controlled access to commercial AI models instead of public chatbots.

  • Reported model tiers: locally hosted open-weight models (Meta's Llama, OpenAI's GPT-OSS, Mistral Small) and externally hosted services (OpenAI's ChatGPT, Anthropic's Claude Sonnet), chosen by task sensitivity.

  • Politico reported a possible September rollout, citing sources; the institution has not formally confirmed that timing.

  • The trigger was roughly 2,100 staff, about a fifth of the total, using AI daily, with committee chairs flagging AI-drafted amendments and invented legal references.

  • The Parliament's press office says the model lineup will be reviewed on European sovereignty grounds; a gateway addresses data governance, not model accuracy.