Source: POLITICO

The European Parliament is testing an internal AI platform called EPGenAI Hub to provide MEPs and staff controlled access to models from Meta, OpenAI, Anthropic and France’s Mistral as an alternative to public generative tools. A POLITICO-reviewed beta lets users choose locally hosted models (Meta’s Llama, OpenAI’s GPT-OSS and Mistral Small) or externally hosted services (OpenAI’s ChatGPT and Anthropic’s Claude Sonnet), and two people familiar with the plans said the rollout could begin as early as September. Parliament leaders want to discourage feeding parliamentary material into uncontrolled public platforms and are worried that AI is degrading legislative quality by producing errors and misleading content.

Committee chairs raised alarms in early July about growing use of AI for drafting amendments, speeches and questions, citing telltale drafting styles, increased translation needs and occasional hallucinated references to non-existent laws. Secretary-General Alessandro Chiocchetti was asked to draft guidance for lawmakers but indicated the issue might fall to the constitutional affairs committee, which would need to amend rules of procedure. Around 2,100 people — about 20% of staff — reportedly use AI daily; staff are covered by April 2024 internal guidelines requiring an “AI-assisted” disclaimer and prohibiting entry of sensitive information into public generative systems. EPGenAI Hub has been presented at workshops promising practical applications, and the Parliament’s press service said model offerings are under revision as open-source and EU sovereign models from sovereign cloud providers become available. The piece notes the tension between opening to U.S. AI companies and the Parliament’s digital-sovereignty moves (such as replacing Google with Qwant and disabling built-in AI features on tablets) and does not report any final institutional decision changing MEPs’ disclosure obligations.