How Parliaments Are Learning to Live With AI
Inside three legislatures’ very different bets on what happens when members and staff start using AI to do their jobs
BY BEATRIZ REY
Legislators and their staff are turning to GenAI tools such as ChatGPT and Claude to draft speeches, bills, and amendments. Substack recently added a scan tool that flags whether a post was written fully, partially, or not at all with AI help. How and when and how much to use AI is now a mainstream question for us all.
And legislatures are responding. Three chambers recently announced changes:
The European Parliament plans to launch its own internal platform, EPGenAIHub, in September,
Chile’s Chamber of Deputies, approved a new internal AI policy last week, and
The US House of Representatives is developing a House AI Center sandbox for experimentation.
Side by side, the three initiatives read like three different theories of how an institution should handle AI. The European Parliament is trying to centralize access through its own platform. Chile is regulating staff use of tools. And the US House is experimenting with something else entirely: treating staff as builders, not just users.
Here’s what I’ve learned about each one.
EPGenAIHub
A staffer in the European Parliament told me: AI is both “a blessing and a curse.” That tension is driving EPGenAIHub, the parliament’s new internal AI platform, set to give lawmakers and staff access to models from Meta, OpenAI, Anthropic, and France’s Mistral.
The staffer explained that staff are already using AI, but largely on their own, through whichever commercial tools they choose. EPGenAIHub is an attempt to give the institution more control over how AI is used internally, while building its own expertise.
The staffer explained that an internal environment gives parliament a say in where information goes, how models are accessed, and which models are used. Eventually, it could also be adapted to parliament’s own material and workflows. The hope is that a tool designed around EU policymaking would be more useful (and more reliable on EU-specific questions) than a generic commercial chatbot.
The initiative sits within a broader political push for the EU parliament to become more institutionally active on AI. The staffer mentioned proposals for AI-focused capacity within the European Parliament’s Secretariat and described a general effort to “reshape” the institution’s organizational structure around the technology.
The platform is also a response to the new reality that generative AI is increasing the workload for legislative staff by multiplying the volume of text moving through the institution: more amendments, more speeches, more material to process (this point also came up recently in POPVOX Foundation’s latest report, State Legislative Capacity in the AI Era). The staffer keeps returning to a “blessing and curse” framing. What worries him isn’t AI itself, but “slop”: drafting that gets faster and shallower at the same time, and an expectation that staff should simply handle more, just because the tool now makes more possible.
Chile’s Chamber of Deputies AI policy
In what I think is a global first, the Chilean Chamber of Deputies has approved a new policy laying out rules, prohibited uses, reporting mechanisms and a new status designation for AI-assisted or drafted documents.
Every document produced for institutional purposes will now carry one of four color-coded declarations in the document or its metadata:
no AI use,
partial use,
predominant or complete use, or
fully automated output with no human involvement at all.
The fourth category comes with a hard restriction: documents produced solely by AI have no institutional value and can never be cited as official sources. The policy goes beyond disclosure to draw a line past which the output stops counting as an institutional document at all. And when staff isn’t sure which label applies, the rule defaults upward, toward the more cautious declaration.
The protocol also sorts tasks by risk:
Public information, routine summaries, brainstorming: low risk, just check for accuracy.
Drafting internal documents: moderate risk, review against reliable sources.
Legal interpretation, legislative advice, budget matters, disciplinary or sanctioning resolutions, personnel decisions: high risk, where AI output is explicitly preliminary, subject to full human validation, and can never be the sole basis for a decision.
Under the new policy, staff have to declare not just whether a given document used AI, but how often they use it in general (occasional, frequent, systematic, or structural). The chamber wants to understand how embedded AI has become in the everyday work of the chamber.
The policy states: “AI use never excuses, reduces, or transfers responsibility.” Every institutional document needs a formally identifiable human author and the human is responsible for its content.
US House Sandbox
The US House approach is taking a very new direction much more akin to an internal startup incubator: a pipeline that lets staff build their own AI tools, with House AI Center engineers shadowing them while they do it.
While Chile’s new policy focuses on what staff cannot do, the US House is inviting staff to build things themselves, with technical experts mentoring them through the process rather than checking their work after the fact. It’s less “use this tool carefully” and more “come build the tool.”
I called these theories at the beginning. Each one hypothesizes something different about where AI actually goes wrong:
the European Parliament thinks the risk is losing control of infrastructure,
Chile thinks it’s losing track of who’s accountable,
the US House thinks it’s staff who can’t keep up.
None of the three has been tested long enough to say which one is right.
Modern Parliament (“ModParl”) is a newsletter from POPVOX Foundation that provides insights into the evolution of legislative institutions worldwide. Learn more and subscribe at modparl.substack.com.
