SerpApi.Org
About 2,100 European Parliament staff — roughly 20 percent of its total workforce — have spent years pasting draft amendments, speeches, and parliamentary questions into public chatbots, with no institutional guardrail stopping them. The results surfaced in committee last month: a surge of amendments arriving for translation (a telltale sign they were composed in AI-generated English), legislative text bearing the distinctive cadence of large language models, and, most damaging, citations to laws that do not exist. The Parliament's answer, reported by Politico on July 20 and confirmed by multiple outlets, is EPGenAI Hub — an internal, sensitivity-tiered AI platform for MEPs and staff, expected to roll out as early as September 2026. There is one catch: when MEPs use it, they still will not be required to say so.
That silence sits at the center of a contradiction that European Parliament watchers have noted but no prior coverage has foregrounded. The Parliament is the institution that passed the AI Act — the world's first binding AI governance framework — and specifically the body whose Article 50 transparency obligations require deployers to disclose when AI has generated content for purposes of informing the public on matters of public interest. Those obligations become enforceable on August 2, 2026, eleven days from today. There is no equivalent obligation binding MEPs themselves when they use AI to draft the laws governing everyone else.
Sensitivity Tiers, Not a Single Chatbot
EPGenAI Hub is not a single AI assistant. Its architecture is designed around a judgment that different legislative tasks carry different data-risk profiles.
The more sensitive the task, the more the model stays close to home. Three AI models run directly on Parliament-controlled servers: Meta's open-weight Llama, OpenAI's GPT-OSS, and Mistral Small from the French AI company of the same name. A draft amendment citing real legislative references, a classified staff memo, a speech draft touching on sensitive negotiations — these tasks are designed to stay within Parliament's own infrastructure, away from commercial cloud environments.
For less sensitive or more computationally demanding work, the platform routes through external providers via managed API connections: OpenAI's ChatGPT and Anthropic's Claude Sonnet. MEPs select the model based on the task's sensitivity level — a design borrowed from enterprise AI governance playbooks now proliferating across regulated industries.
The full model lineup is still described by the Parliament's press service as under review. Open-source and EU-origin sovereign models are increasingly available through cloud providers recently contracted by the European Commission, and may appear in future iterations of the hub.
Staff were introduced to the platform through internal workshops, including one promoted as a way to help employees work smarter and faster. Beta testing is ongoing.
Lawmakers Set Rules Others Must Follow — and Exempt Themselves
The governance gap behind EPGenAI Hub is harder to paper over than the platform itself. EP staff have operated under AI use guidelines since April 2024, requiring them to label AI-assisted work and prohibiting them from feeding sensitive information into public AI tools. MEPs face no equivalent obligation.
As of today, there is no rule requiring a Member of the European Parliament to disclose whether an amendment, a parliamentary question, or a plenary speech was drafted, in whole or in part, by an AI system. When committee chairs raised the alarm over AI quality failures earlier this month and asked the secretary-general to develop guidelines, the response was that new MEP rules may require the Constitutional Affairs Committee to rewrite Parliament's standing rules entirely — a multi-year process, not an executive memo.
Brando Benifei, the Italian MEP who co-authored the EU AI Act and steered it through Parliament, named the problem directly. Speaking to The Next Web, Benifei said: "Parliament has not been immune to AI slop and misleading AI-generated content. MEPs must lead by example as the AI Act's transparency rules become applicable."
His fellow MEP Barry Andrews, an Irish Renew lawmaker, acknowledged that AI use is plainly visible across Parliament's legislative output. "I fully understand, given our workload," Andrews said of colleagues who rely on AI to draft questions, speeches, and amendments. "But for me personally, these are core democratic functions which voters elected us to perform."
The optics are pointed. The Parliament is now equipping itself with AI tools just as the AI Act's disclosure requirements — the rules it designed for the rest of the world — are about to become enforceable.
What Article 50 Actually Requires — and Why the Parliament Is in a Complicated Position
Article 50 of Regulation (EU) 2024/1689 creates four categories of transparency obligation, enforceable from August 2, 2026. The one directly relevant to legislative work is Article 50(4): deployers who use AI systems to generate text that is published on matters of public interest, for the purpose of informing the public, must disclose that the content is AI-generated.
Whether that provision binds the Parliament's internal operations — as opposed to its published communications — is not straightforward. EU institutions occupy a specific legal position in relation to EU market regulations, and Regulation 2024/1689 was designed primarily for market actors rather than the Parliament acting in its legislative capacity. But the broader question of what "leading by example" means is a political one, not just a legal one. If the Parliament publishes AI-assisted position papers, public explanations of legislation, or institutional communications, Article 50's requirements apply. If a MEP uses an AI model to draft a legislative amendment and that amendment enters the official parliamentary record, the transparency question is live regardless of whether Article 50 technically compels an answer.
The EU Commission, for its part, published final guidelines on Article 50 compliance on July 20 — the same day the EPGenAI Hub story broke — confirming that the transparency obligations are not delayed and will apply across all AI systems within their scope from August 2.
Does Deploying US AI Undercut European Digital Sovereignty?
The EPGenAI Hub launch sits awkwardly against the Parliament's own stated trajectory on digital sovereignty. In June 2026, Parliament's IT team rewired thousands of computers so that searches now route through Qwant, the French privacy-focused search engine, instead of Google — a pointed signal that the institution wants to reduce its dependence on American digital infrastructure. Parliament also previously disabled built-in AI assistants on MEPs' tablets over data-protection concerns.
EPGenAI Hub's model lineup reads differently: OpenAI, Meta, Anthropic. Three of the five AI providers in the hub are US-headquartered, with the largest share of anticipated usage likely flowing to the most capable external models — ChatGPT and Claude Sonnet — rather than the on-premise alternatives.
The tension is real, and Parliament has not resolved it. The press service's acknowledgment that the model roster is under review suggests that European and open-source alternatives could displace some or all of the US providers in future versions. The European Commission has contracted cloud providers offering EU-origin sovereign AI models, and Parliament has signaled that those options are on the table. Whether commercial capability gaps between current EU-origin models and frontier US models will prove durable enough to make that substitution workable is a separate question the hub's rollout will eventually force an answer to.
What Comes Next for Enterprise AI Teams
The European Parliament's experience is arriving in front of enterprise IT and policy professionals at precisely the moment when most large institutions are confronting the same shadow IT problem in their own AI ecosystems. The pattern — uncontrolled adoption at scale, followed by a quality failure, followed by a sanctioned alternative that tries to route usage into a governed environment — is repeating across sectors.
The specific architecture EPGenAI Hub employs (sensitivity tiering, on-premise for high-risk workloads, managed API access for commodity tasks, multi-model flexibility) is increasingly the recognized institutional response. It is not a novel design. What is specific to the Parliament is the institutional weight of the accountability question: an institution whose core output is authoritative legal text, and whose hallucination problem is literally a falsification of the legal record, faces stakes that most enterprise deployments do not.
The September rollout, if it holds, would deliver the hub to MEPs at almost exactly the moment Article 50 begins to bind the rest of the organizations in the European economy. Whether Parliament uses that timing to set a standard — by establishing a voluntary disclosure norm, even without a formal rule — or simply to give MEPs a better tool without governance obligations attached, will be a meaningful early indicator of what "leading by example" actually means in practice.
Frequently Asked Questions
Are MEPs required to disclose whether they used AI to write legislation?
No. As of today, no rule requires a Member of the European Parliament to disclose AI-assisted drafting of amendments, parliamentary questions, or speeches. EP staff have operated under voluntary guidelines since April 2024 requiring them to label AI-assisted work, but MEPs are not covered. Creating an equivalent MEP obligation would require the Constitutional Affairs Committee to rewrite Parliament's standing rules — a significantly longer process than a policy memo. The launch of EPGenAI Hub does not, on its own, create any disclosure obligation.
What is EPGenAI Hub and which AI models does it include?
EPGenAI Hub is the European Parliament's new internal AI platform, currently in beta testing and expected to roll out to MEPs and staff in September 2026. It gives users access to a range of AI models chosen based on task sensitivity. On-premise models running on Parliament's own servers include Meta's Llama, OpenAI's GPT-OSS, and Mistral Small. External models accessible via managed API connections include OpenAI's ChatGPT and Anthropic's Claude Sonnet. The full roster is described by Parliament's press service as still under review, with EU-sovereign and open-source alternatives under consideration.
When does the EU AI Act's Article 50 transparency requirement take effect?
Article 50 of Regulation (EU) 2024/1689 becomes enforceable on August 2, 2026 — eleven days from today. It requires, among other things, that deployers who use AI systems to produce content on matters of public interest, for the purpose of informing the public, disclose that the content is AI-generated. Organizations that deployed generative AI systems before August 2 receive a grace period until December 2, 2026, for the specific watermarking requirement under Article 50(2). Violations carry fines of up to €15 million or three percent of global annual turnover.
Does the European Parliament's use of AI through EPGenAI Hub comply with Article 50?
This is unsettled. EU institutions occupy a complex position in relation to EU market regulations, which were designed primarily for private-sector actors. Whether Article 50 directly binds the Parliament's internal legislative operations is a legal question that has not been formally resolved. What is clear is that if Parliament publishes AI-generated content on matters of public interest — such as institutional communications or public policy explanations — Article 50's disclosure requirements would apply. The political question of whether MEPs should voluntarily disclose AI-drafted legislative work, regardless of legal compulsion, is what Benifei and other MEPs are now grappling with.
Related Articles
