Trustworthiness in Generative Foundation Models Is Still Poorly Understood

TMLR Paper5158 Authors

Abstract: Generative Foundation Models (GenFMs) have seen extensive deployment across diverse domains, significantly impacting society yet simultaneously raising critical concerns about their trustworthiness, including misinformation, safety risks, fairness, and privacy violations. Recognizing the complex nature of these issues, to bridge the gap between abstract principles and operational actions throughout the GenFM lifecycle, we propose a flexible and multidimensional set of trustworthiness guidelines. These guidelines incorporate ethical principles, legal standards, and operational needs, addressing key dimensions such as fairness, transparency, human oversight, accountability, robustness, harmlessness, truthfulness, and privacy. Our guidelines serve as adaptable tools to bridge abstract principles and practical implementations across varied scenarios. Building upon these guidelines, we identify several core challenges currently unresolved in both theory and practice. Specifically, we examine the dynamic tension between adaptability and consistent safety, the ambiguities in defining and detecting harmful content, and the balancing of trustworthiness with model utility. Through our analysis, we reveal that the trustworthiness of GenFMs remains inadequately understood, highlighting the necessity for continuous, context-sensitive evaluation approaches. Consequently, we propose potential solutions and methodological directions, emphasizing integrated strategies that combine internal alignment mechanisms with external safeguards. Our findings underscore that trustworthiness is not static but rather demands ongoing refinement to ensure the responsible, fair, and safe deployment of GenFMs across various application domains.

Submission Length: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Jaakko_Peltonen1

Submission Number: 5158