Formalizing FOIA Exemption Standards as Computable Legal Norms: A Framework for LLM-Assisted Sensitivity Review

Romana Afroze

Keywords: FOIA, sensitivity review, computable legal norms, deontic logic, large language models, explainability, open government, access to information

TL;DR: The paper proposes a framework using Legal Norm Ontology and NIPA to improve FOIA sensitivity reviews by treating exemptions as computable legal norms rather than simple classification tasks.

Abstract:

Governments worldwide are deploying artificial intelligence to assist with the Freedom of Information Act (FOIA) and Access to Information
(ATI) review, yet existing systems treat sensitivity classification as a pure machine-learning task, ignoring the normative legal structure encoded in the exemption doctrine. This paper argues that FOIA exemptions are not labels but computable legal norms: structured balancing tests whose elements can be formalized using deontic logic and ontological modelling. This paper proposes a two-layer framework—a Legal Norm Ontology (LNO) for the nine FOIA exemptions and a Norm-Informed Prompting Architecture (NIPA) that operationalizes the ontology within large language model (LLM) pipelines. NIPA is evaluated against two baselines on a curated corpus of 847 judicially reviewed FOIA redaction decisions drawn from D.C. Circuit litigation records, demonstrating a 14.4-percentage-point improvement in exemption classification precision over the fine-tuned baseline, and a 22.1-percentage-point gain in legal reasoning alignment (LRA—the proportion of criterion-level assessments whose rationale matches the ground-truth annotation protocol) over the zero-shot baseline. A Disclosure-of-Reasoning (DoR) standard is derived from the administrative due process doctrine and argued to constitute a necessary condition for lawful deployment—not merely a governance aspiration. These findings have direct implications for AI governance frameworks in open government contexts globally.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 10