Oct 27, 2024
Mapping Intent: Documenting Policy Adherence with Ontology Extraction
Alejandra de Brunner, Mia Hopman, Jack Wittmayer
This project addresses the AI policy challenge of governing agentic systems by making their decision-making processes more accessible. Our solution utilizes an adaptive policy ontology integrated into a chatbot to clearly visualize and analyze its decision-making process. By creating explicit mappings between user inputs, policy rules, and risk levels, our system enables better governance of AI agents by making their reasoning traceable and adjustable. This approach facilitates continuous policy refinement and could aid in detecting and mitigating harmful outcomes. Our results demonstrate this with the example of “tricking” an agent into giving violent advice by caveating the request saying it is for a “video game”. Indeed, the ontology clearly shows where the policy falls short. This approach could be scaled to provide more interpretable documentation of AI chatbot conversations, which policy advisers could directly access to inform their specifications.
William Jurayj
Love the use the visualization graph!
Jaime Raldua
Great work! I could see policy makers using an upgraded version of this tool.
Abe Hou
This is an interesting visualization and idea. Very nice and innovative ideas. Would appreciate if you justify how well this can be applied in real policy settings, because it seems that it doesn't necessarily align with actual policy tasks
Cite this work
@misc {
title={
@misc {
},
author={
Alejandra de Brunner, Mia Hopman, Jack Wittmayer
},
date={
10/27/24
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}