Topics: AI, compliance, natural language, evaluation, plain language
Technologies: python, LLM APIs, prompt design, evaluation methods
Tldr; build and rigorously evaluate the AI-assisted checks in a civic compliance checker, for questions like ‘is this privacy policy actually written in plain language’, where a pattern match cannot give the answer.
Motivation: Some compliance questions are irreducibly qualitative. Whether a privacy policy is genuinely readable, whether a complaints process is clearly explained, whether funding is disclosed in an intelligible way: these cannot be answered by matching against a list. A compliance checker can use a language model to make such judgements, but this raises its own problem. An AI-assisted check that is confidently wrong is more dangerous than no check at all, because it launders a guess into an apparent verdict. This project builds a small set of AI-assisted checks and, more importantly, evaluates how trustworthy they are, establishing where a model’s judgement can be relied upon and where a human must remain in the loop.
Implementation: You will identify civic standard requirements that resist deterministic checking, such as plain-language obligations under the Accessibility and Transparency standards. You will design AI-assisted checks for them, giving careful attention to prompt design, to how the model is asked to justify its verdict, and to how uncertain cases are surfaced rather than hidden. You will then build a labelled evaluation set, with human judgements on the same material, and measure how far the model agrees with human assessors, where it diverges, and whether it fails safely (flagging for review) or unsafely (asserting a wrong verdict confidently). The deliverable is a working set of AI-assisted checks and a candid evaluation of their reliability, including recommendations on where human oversight is mandatory.
What you will learn: (i) which compliance questions genuinely require judgement rather than matching; (ii) how to design and constrain LLM-based checks; (iii) how to evaluate a model’s agreement with human assessors; (iv) how to build human oversight into an automated system; (v) the risks of automating qualitative judgement in a high-stakes setting.
References:
[1] dotPublic compliance checker prototype: https://dotpublic.org/checker
[2] Web Content Accessibility Guidelines (WCAG): https://www.w3.org/WAI/standards-guidelines/wcag/