Scholarlink

Context and motivation: The increasing adoption of AI in public administration requires translating ethical and legal expectations into verifiable requirements. In Brazil, AI is increasingly explored to support eligibility assessment and allocation in social programs, yet there is limited guidance on how to specify and validate fairness and transparency in such high-stakes settings. Question/problem: This research preview investigates how fairness and transparency can be operationalized as measurable non-functional requirements (NFRs) for AI-supported social benefit allocation. Principal ideas/results: As a proof-of-concept, we simulate scholarship allocation using ProUni data and assess the model through predictive performance, group fairness metrics, and SHAP-based explainability evidence. Results indicate satisfactory accuracy while revealing measurable racial disparities, motivating bias mitigation. Contribution: We propose an initial RE-oriented framework that integrates fairness indicators, explainability artifacts, and accountability mechanisms into specification and validation checkpoints, supporting early-stage discussion and feedback on responsible AI in digital government.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Specifying Fairness and Transparency Requirements for Public Benefit Allocation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Amanda Aline F. C. Vicenzi,
  • José Siqueira de Cerqueira,
  • Pekka Abrahamsson,
  • Edna Dias Canedo

摘要

Context and motivation: The increasing adoption of AI in public administration requires translating ethical and legal expectations into verifiable requirements. In Brazil, AI is increasingly explored to support eligibility assessment and allocation in social programs, yet there is limited guidance on how to specify and validate fairness and transparency in such high-stakes settings. Question/problem: This research preview investigates how fairness and transparency can be operationalized as measurable non-functional requirements (NFRs) for AI-supported social benefit allocation. Principal ideas/results: As a proof-of-concept, we simulate scholarship allocation using ProUni data and assess the model through predictive performance, group fairness metrics, and SHAP-based explainability evidence. Results indicate satisfactory accuracy while revealing measurable racial disparities, motivating bias mitigation. Contribution: We propose an initial RE-oriented framework that integrates fairness indicators, explainability artifacts, and accountability mechanisms into specification and validation checkpoints, supporting early-stage discussion and feedback on responsible AI in digital government.