Simulating Policy Discussions with Large Language Models
This project explores how large language models (LLMs) can simulate parliamentary policy discussions using digital traces from the UK Parliament’s Hansard debates, providing a new way to study and experiment with complex policymaking processes.
Parliamentary debates provide rich evidence of democratic decision-making, capturing diverse perspectives from policymakers and stakeholders. However, analysing these large-scale conversational records remains challenging. This project investigates whether large language models (LLMs) can simulate UK parliamentary debates by learning from digital footprints contained in Hansard transcripts.
The study compares simulated debates generated by different LLMs with real parliamentary discussions. By examining speaker roles, party affiliations, and stance-taking behaviours, the project evaluates how closely AI-generated debates reflect real policy dialogue dynamics. In the longer term, this research aims to support the development of a “deliberation sandbox”, a digital environment where complex policy discussions involving multiple stakeholders can be simulated and explored.
The study uses transcripts from Hansard as digital footprints representing real parliamentary discussions. Multiple large language models, including Gemini 2.5, ChatGPT‑o3, and GovernmentGPT (an open-source model fine-tuned on Hansard), were prompted to generate simulated parliamentary debates on the test topic of ultra-processed foods.
Each generated debate included speakers with party affiliations and corresponding speech content. The simulated discussions were then compared with real parliamentary debates to evaluate similarities in participation patterns, party representation, and stance-taking behaviour.
Initial findings suggest that LLM-generated debates can reproduce several structural features of real parliamentary discussions. In particular, the distribution of speakers by political party in the simulated debates broadly reflects the composition of actual parliamentary debates.
However, many generated statements show neutral or ambiguous stances, indicating challenges for current models in producing clear argumentative positions. These findings highlight both the potential and limitations of using LLMs to model complex policy dialogues.
Simulating Policy Discussions with Digital Footprints and Large Language Models.
Parliaments play a significant role in democratic decision-making, drawing on evidence from a wide range of stakeholders. This evidence data, published on official government websites, is a valuable archive of digital traces reflecting… [more]
Link to the Digital Footprints 2025 Conference Award-Winning Poster.
