Nov 2, 2025
Modeling the political process to forecast the outcomes of hypothetical AI governance proposals
Linh Le, David Williams-King
International cooperation on AI governance faces a fundamental trust problem: countries like the US and China struggle to assess whether proposed agreements would actually be implemented by their counterpart's domestic political systems. This uncertainty undermines the credibility of commitments and hinders the establishment of safety regulations needed to prevent catastrophic AI risks. We address this challenge by developing a system that predicts whether AI safety legislation would gain support within a country's government, enabling both domestic policymakers and international partners to evaluate the political feasibility of proposed regulations. Our approach uses large language models to generate interpretable yes/no questions about legislative bills, then learns legislator-specific perspective representations that capture individual voting patterns on AI policy. We collect and analyze voting records from the 118th and 119th U.S. Congresses (2024-2025), identifying 146 AI safety-related bills. Our model significantly outperforms baseline approaches in forecasting senatorial votes on AI legislation. Additionally, we develop a suite of hypothetical AI governance policies ranging from strict to permissive, using our model to identify political feasibility thresholds—the boundaries between policies likely to pass versus fail. This work provides a concrete tool for improving transparency and trust in international AI governance negotiations.
A novel and an exciting approach - I’d be excited to see you work on this more!
However, I’m quite confused by how the accuracy of the model is measured. The way I understand this works now is:
-When training, the model learns from the sponsorship patterns
-In testing, you ask whether a senator that is mentioned in the bill would support it. This feels odd to me - doesn’t the fact that a senator is mentioned in the bill mean that they support the bill?
It would be beneficial to hold out entire bills from training and then test accuracy on those, so that you can measure accuracy more accurately.
Currently, there’s no validation for the main use case. You generate predictions for the gradated hypothetical policies, but for them, there is no ground truth.
I also think the baseline comparison is flawed - GPT3-oss is not a state-of-the-art AI forecasting tool (you could have compared to this, for example: https://safe.ai/blog/forecasting). And the cutoff date makes the comparison quite unfair.
Cite this work
@misc {
title={
(HckPrj) Modeling the political process to forecast the outcomes of hypothetical AI governance proposals
},
author={
Linh Le, David Williams-King
},
date={
11/2/25
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
