Brussels lawmakers have a problem. Roughly one in five staff members in the European Parliament already tap public artificial intelligence services for help drafting speeches, questions and amendments. The results have not always inspired confidence. Misleading content generated by these systems has slipped into official materials. Now the institution that passed the world’s first comprehensive AI law is fighting fire with fire.
It calls the project EPGenAI Hub. The secure platform will bundle access to models from OpenAI, Meta, Anthropic and Mistral. Lawmakers and their aides will query them inside parliament-controlled infrastructure or approved external hosting. No more feeding sensitive legislative drafts into ChatGPT on personal laptops. The system, already demonstrated in workshops this week, could go live as soon as September.
But the move raises fresh questions. If the goal is to reduce errors and preserve democratic accountability, why invite the same companies whose products created the mess? And will voluntary disclaimers really stop sloppy AI output from shaping European policy?
Brando Benifei saw the risk early. The Italian Socialist and Democrats lawmaker steered the AI Act through parliament. He did not mince words. “Parliament has not been immune to AI slop and misleading AI-generated content,” he told colleagues. MEPs must lead by example as the AI Act’s transparency rules kick in. The quote appeared first in a TechRadar report published Monday.
Benifei’s warning carries weight. The AI Act, set to apply in stages through 2027, demands clear labeling of AI-generated content in many contexts. Yet inside parliament itself, disclosure remains patchy. Guidelines urge staff to mark modified documents “AI-assisted.” Mandatory rules for every draft do not exist. That gap matters when amendments can alter trade deals or climate targets.
Barry Andrews takes a more relaxed view. The Irish Renew member chairs the Development Committee. He acknowledges colleagues lean on AI to handle crushing workloads. “That is their choice,” Andrews said. “But for me personally, these are core democratic functions which voters elected us to perform.” His comment, also captured by TechRadar, hints at a split. Some see efficiency. Others see erosion of elected responsibility.
The parliament is not alone in its scramble. Recent hearings exposed tensions between EU ambitions and American AI providers. Anthropic sent a junior policy staffer to testify remotely before a key committee instead of its head of public policy. One MEP reacted sharply. The episode, detailed in a July 15 Politico article, underscored worries about sovereignty. European institutions want frontier models without ceding control.
So they built their own gate. EPGenAI Hub runs on parliament infrastructure with options for external hosting that meets security standards. Users cannot pipe classified data into public endpoints. The setup promises to keep conversations inside approved boundaries while still delivering the speed of Claude, GPT or Mistral’s latest releases. Workshops promised “practical, everyday applications that will help you work smarter and faster,” according to an internal email reviewed by Politico.
Yet technical safeguards only go so far. Generative models still hallucinate. They invent facts, citations, even entire legal precedents when pressed. Parliamentary researchers already caught such fabrications in early experiments. The new platform may reduce leakage of sensitive information. It does not magically eliminate the models’ tendency to sound authoritative while being wrong.
Critics argue the parliament is simply outsourcing judgment. By offering sanctioned access, it normalizes AI assistance in legislative drafting. That shift carries long-term consequences. Constituents expect human deliberation on issues from migration to subsidies. When an assistant prompts Claude to generate a 15-page amendment, who really authored the text?
Transparency offers one defense. The AI Act requires providers to document training data and enable detection of synthetic content. Parliament could demand those tools be integrated into EPGenAI Hub. Watermarks, metadata logs, mandatory source citations. None of those steps appear confirmed in current plans. And even perfect labeling cannot fix flawed reasoning.
Recent coverage highlights the speed of adoption. A story from The Next Web on Monday noted the platform gives MEPs “sanctioned access” precisely to stop unsanctioned use. Another report from Cybernews emphasized the internal focus on curbing public tool reliance. The coverage, all published within the past 24 hours, shows how quickly the story spread after initial leaks.
Numbers tell part of the story. Parliament employs roughly 10,500 people. About 2,100 already experiment with AI, per internal surveys cited in the TechRadar piece. That figure likely undercounts actual usage. Assistants juggle multiple roles and tight deadlines. The temptation to let an LLM summarize a 200-page impact assessment proves hard to resist.
Supporters counter that prohibition would fail. Better to channel the technology into monitored channels than pretend staff will ignore tools available on every smartphone. The hub, they say, lets parliament set acceptable use policies, monitor prompts for sensitive content, and perhaps fine-tune models on EU legal texts. Mistral’s European base adds a veneer of digital sovereignty that pure American offerings lack.
Still, the reliance on those same American firms stands out. OpenAI, Anthropic and Meta dominate frontier capabilities. European alternatives trail. The parliament’s choice reflects market reality even as officials preach strategic autonomy. One committee hearing after another hammers home the point: access to compute, data and talent tilts across the Atlantic.
Implementation details remain sparse. Testing continues. Rollout timing could slip. Integration with existing document systems, audit trails, user training. All demand attention before September. And larger questions loom. Will MEPs publicly disclose when they used the hub on a bill? Will committees adopt standards for AI-generated evidence? The AI Act itself leaves some of these judgments to member states and institutions.
Benifei believes parliament must model compliance. His work on the legislation spanned years of negotiation. He understands both the promise and the pitfalls. If the hub succeeds, it could become a template for other EU bodies. National assemblies in Berlin, Paris or Warsaw might copy the approach. Failure would hand ammunition to those who argue the AI Act cannot be enforced even by its creators.
Andrews’ skepticism feels grounded too. Democracy runs on human accountability. Voters elect individuals, not algorithms. When an MEP stands in plenary to defend an amendment, the public expects that the arguments, facts and compromises reflect personal judgment. AI can assist research. It should not substitute for it.
The tension will not resolve quickly. Generative models grow more powerful each quarter. Costs fall. Capabilities expand. Parliament’s modest hub may look quaint in two years when multimodal systems ingest entire legislative histories and spit out ready-to-vote packages. The institution is buying time. Whether it uses that time to build genuine European alternatives or simply becomes more dependent remains the open bet.
One thing seems clear. The era of pretending AI stays outside the chamber has ended. Lawmakers who once regulated from a distance now sit inside the machine. Their experiment will be watched closely. Not just in Brussels, but in every capital wrestling with the same seductive, error-prone tools.
WebProNews is an iEntry Publication