Collaborative Research: DASS: Policy Design for Holding AI-Supported Systems Accountable

随着社会对软件系统的依赖程度越来越高，如何负责任地开发这样的软件就产生了问题。在设计软件系统时，必须在灵活性和创新性与保护公众利益之间取得平衡。制定管理软件对社会影响的政策的监管机构往往很难获得明确的指导，以追究软件系统的责任，并跟上新技术的发展，例如人工智能(AI)的日益使用。软件工程师往往只能得到模糊的指导，也没有可衡量的目标，比如欧盟的“解释权”。虽然澄清指导意见或法院裁决最终可能提供更多可操作的细节，但政策执行是一个缓慢的反应过程，可能会推迟改进措施的采用。为了协调软件和策略设计，并为负责任的软件系统构建主动的策略框架，需要更多地关注可解释性、安全性和公平性等质量指导。本项目将调查策略设计的跨域原则，为软件工程师设定相应的义务，特别是在人工智能支持的系统的背景下。结果将帮助利益相关者有意识地为特定领域设计基于证据的指导政策。该项目将促进政策制定者和软件工程师之间的互动，以了解形成政策目标的能力和社会期望，调查政策维度和涉及的具体权衡，并将政策与具体的质量保证义务联系起来，这些义务可用于涵盖系统人工智能和非人工智能部分的监管评估。具体地说，该项目将首先使用利益相关者访谈、历史分析和技术分析来调查政策设计的重要维度(例如，监管的严格性、证据的水平、政策的特殊性)。然后，它将通过几个实验来评估设计决策如何影响结果、相互作用，以及如何产生权衡--这些权衡将被编码和验证为策略设计模式。本研究将结合多种研究方法，包括访谈法、文献分析法、对照实验法、样本法和作者工作坊。它将把关于政策设计和监管的社会科学研究与关于系统设计和质量保证的软件工程研究深度结合起来。这项研究将通过与监管机构的接触，在多个社区广泛传播。该奖项反映了NSF的法定使命，并通过使用基金会的学术价值和更广泛的影响审查标准进行评估，被认为值得支持。

As society is becoming more and more dependent on software systems, questions arise about how such software can be developed responsibly. In the design of software systems, flexibility and innovation must be balanced with safeguarding the public. Regulators who design policy to steward software’s effect on society often have a hard time capturing clear guidance to hold software systems accountable and keeping up with new technologies, such as the increased use of artificial intelligence (AI). Software engineers have often been left with vague guidance and no measurable goals, such as the EU’s “right to explanation.” While clarifying guidance or court decisions may eventually provide more actionable details, policy implementation is a slow, reactive process that can delay adoption of improvements. To reconcile software and policy design and construct a proactive policy framework for accountable software systems, guidance for qualities such as explainability, safety, and fairness requires more attention.This project will investigate cross-domain principles for policy design, setting corresponding obligations for software engineers, particularly in the context of AI-supported systems. Results will help stakeholders to deliberately design policy with evidence-based guidance for a specific domain. The project will facilitate interactions between policy makers and software engineers in understanding capabilities and societal expectations toward forming policy goals, investigate policy dimensions and the concrete tradeoffs involved, and connect the policy to specific quality-assurance obligations that could be used in a regulatory evaluation covering both AI and non-AI parts of the system. Specifically, the project will first investigate the important dimensions for policy design (e.g., strictness of regulation, level of evidence, policy specificity) using stakeholder interviews, historical analysis, and technical analysis. It then will evaluate, with several experiments, how design decisions affect outcomes, interact with each other, and result in tradeoffs -- which will be encoded and validated as policy design patterns. The research will combine multiple research methods, including interviews, literature and document analyses, controlled experiments, prototyping, and writer’s workshops. It will deeply integrate social-science research on policy design and regulation with software-engineering research on system design and quality assurance. The research will be disseminated broadly across multiple communities and through engagement with regulators. Educational material will be developed for multiple courses and shared publicly.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.