Who Shapes the First Interpretation? A Fork in the Logic of Legislative Work
Large language models may either concentrate control over how issues are framed or expand committees’ capacity to produce their own analysis
BY BEATRIZ REY
Committee deliberation in Brazil’s Chamber of Deputies (Photo credit: Lula Marques/Agência Brasil)
I recently came across the article “Approaching the integration of large language models in the parliamentary workspace” by Jörn von Lucke, which examines how large language models (LLMs) could be introduced into parliamentary work.
The author argues that LLMs offer clear gains in efficiency (supporting drafting, summarization, and information retrieval), but the real challenge lies in governance. Parliaments must decide whether to rely on external providers, build in-house systems (an option currently pursued by only a small number of parliaments), or develop their own models (likely beyond the reach of most, if not all, parliaments), all while managing risks related to bias, transparency, cybersecurity, and accountability. His core claim is that AI adoption is not primarily a technical issue, but an institutional one: without the right capacity, rules, and oversight, these tools may distort rather than improve legislative decision-making.
Reading it, I kept coming back to something that sits just beneath the surface of this debate: a quiet shift in how legislative work is organized.
For most of modern parliamentary history, legislatures have been structured around a basic constraint: information is limited, unevenly distributed, and costly to process. A significant part of the institution’s design, from committees to research services, exists to manage that scarcity. The problem has been how to gather information, how to make it credible, and how to move it through a system where actors disagree.
Thinking about this now, I see two distinct possibilities for how this constraint interacts with the arrival of AI.
The first is the one that initially motivated this reflection. As AI systems increasingly generate pre-structured outputs, such as summaries, arguments, comparisons, and draft texts, the locus of interpretation shifts upstream. The risk is that part of the interpretive work is done before it enters the visible legislative process.
What changes here is not just efficiency, but the stage at which meaning is constructed. Large language models do not simply retrieve information; they organize it. They can be used to surface what appears salient, suggesting relevant comparisons, and structure how a problem is presented (depending on how they are prompted and configured). Information can arrive not as relatively raw material to be debated, but as already synthesized outputs (when that is the form of output being requested; these outputs reflect the assumptions embedded in the user’s prompt and framing).
In this scenario, the structuring of outputs reflects both how users formulate prompts and how systems are designed, configured, and governed – choices that shape what kinds of outputs are possible – and deliberation becomes, at least in part, a response to these pre-structured representations.
As a result, part of the interpretative work shifts upstream and becomes less visible within the deliberative process itself. When AI-generated outputs appear coherent and neutral, the underlying assumptions can become less visible. The challenge is not simply to keep the human in the loop, but to ensure that the structuring of information by these systems remains open to scrutiny.
But there is a second possibility that becomes visible once we take seriously how constrained committees historically have been.
In practice, committees and their staff have rarely had the time, resources, or technical capacity to engage directly with raw data. Instead, they have depended on external actors, including think tanks, agencies, and, very often, lobbyists, not only to supply information but also to interpret it. As political scientists Richard Hall and Alan Deardorff argue, lobbying often operates less as persuasion and more as a form of legislative subsidy: a transfer of policy information, political intelligence, and legislative labor to allied legislators. In other words, lobbyists help legislators do the work they already intend to do.
Under these conditions, the effective toolkit available to a committee staffer was limited. Much of the work consisted of comparing one well-resourced interpretation against another, one lobbyist’s framing against a competing one. What was largely missing was the capacity to independently interrogate underlying data, test specific use cases, incorporate granular or constituent-level input, or generate original analysis within the institution itself.
From this perspective, AI introduces a different kind of shift. For the first time, there is a plausible path for committees to internalize at least part of that analytical capacity at scale. The ability to query data directly, ask iterative questions, explore edge cases, and synthesize dispersed inputs begins to approximate a form of in-house capability that was previously out of reach.
In that sense, these tools can be understood as an initial (yet still fragile) counterweight to the logic of legislative subsidy. If external actors historically derived influence from supplying scarce analytical labor, expanding internal capacity changes that relationship at the margin.
But this rebalancing is contingent rather than automatic. If lobbyists and other external actors adopt these tools more quickly or deploy them more effectively, the same subsidy dynamic may simply be reproduced at a higher level of sophistication. Instead of competing memos, committees may face competing AI-generated interpretations that remain externally structured, but are now more complex and potentially more difficult to unpack.
The result is not a single trajectory, but a fork. AI can either obscure the origins of interpretation further, or begin to redistribute the capacity to produce it. Which of these paths dominates depends less on the technology itself and more on how, and by whom, it is integrated into legislative work.
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.
