Eager: causal auditing of LLM recommender systems for trustworthy decision support

The National Science Foundation Division of Computing and Communication Foundations awarded Lehigh University $230,000 on September 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) for an EAGER project on causal auditing of large language model recommender systems.

The research develops methods to improve transparency and accountability of LLM-based recommender systems through causal auditing rather than examination of internal model mechanisms. The project reconstructs external explanations from observable recommendation behavior, identifying the purposes of recommendation activities, reconstructing causal influence relationships among information sources and user preferences, and auditing whether recommendations are adequately supported by available information. The work advances fundamental knowledge at the intersection of artificial intelligence, recommender systems, causal reasoning, and knowledge representation, contributing to national efforts to strengthen trustworthy AI technologies and maintain U.S. leadership in AI innovation.

The award supports a two-year period of performance from September 1, 2026, through August 31, 2028, with place of performance in Fountain Hill, Pennsylvania. The assistance type is a Project Grant.

Abstract

Artificial intelligence (AI)-generated recommendations are becoming an increasingly important component of modern digital infrastructure, influencing decisions in e-commerce, workforce development, healthcare, emergency planning, and other domains. An example of this is recommender systems based on large language models (llms), which are rapidly emerging as a powerful new paradigm because they support natural language interaction and flexible decision support. However, these systems often operate as opaque black boxes. As a result, users and organizations may find IT difficult to understand why a recommendation was produced or whether IT is supported by available information, constraints, and preferences. This project investigates a novel research approach for improving the transparency and accountability of LLM-based recommender systems through causal auditing. Rather than examining internal model mechanisms, the research seeks to reconstruct external explanations from observable recommendation behavior. The project will develop methods for identifying the purposes of recommendation activities, reconstructing causal influence relationships among information sources, constraints, and user preferences, and auditing whether recommendations are adequately supported by the information available to the system. The project will advance fundamental knowledge at the intersection of artificial intelligence, recommender systems, causal reasoning, and knowledge representation. By advancing methods for auditing and increasing confidence in AI-generated recommendations, the research contributes to ongoing national efforts to strengthen trustworthy artificial intelligence technologies and maintain U.S. Leadership in AI innovation. The resulting methods aim at improving the transparency and accountability of AI-driven decision support across a wide range of commercial, societal, and public-sector applications. The project will provide training opportunities for graduate students in artificial intelligence, causal reasoning, and trustworthy computing, and research products will be disseminated through publications and public repositories. 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.- subawards are not planned for this award.