Auditing How AI Systems Convert Political Questions Into Technical One

AI systems are already working in civic and governance contexts in the form of content moderation, benefits allocation and resource distribution. Most of these systems are designed to optimise; they take a stated objective and find the most efficient, cost-effective recommendation. This design choice carries the risk of power concentration. These decisions are meant to be made by elected humans; that is their role. They require value trade-offs that need to be contested politically, not resolved in a black-box model with opaque weighting set by engineers, not the public. I call this interpretive opacity: where a system can be technically transparent (open source, auditable) but the interpretive choices that structure it remain invisible.

I built a working prototype (Community Document Analysis tool) that takes the opposite approach. It surfaces the value trade-offs in civic documents rather than trying to optimise. I have tested it against a real Section 21 (UK) eviction notice and it grounds outputs against contestable policy inputs.

If left unaddressed, the risk is a concentration of power in whoever sets the objective function, without any intended bias or accountability for it. As civic infrastructure increasingly adopts optimising AI by default, the risk will compound outside of public scrutiny, offering the public no recourse on the optimised AI recommendations

This grant would allow for the extension of the prototype to a second document type and the write-up of a framework that could be reused as a methodology. The aim of the prototype and the framework is to answer the following: "How do you build AI that mediates instead of optimises?" Aimed at civil society and civic tech practitioners who are currently defaulting to optimiser-style tools without realising the power-concentration cost. I am building this independently, drawing on a decade in regulated AI/software governance.