Assessing AI Certification Validity via Knowledge Graphs for Use Case Similarity

Elizabeth M. Daly, Jasmina Gajcin, Inge Vejsbjerg, Delaram Golpayegani, Rahul Nair, Kush R. Varshney

Keywords: Governance, Knowledge Graphs, AI Certification, AI Use Case

Abstract: With increased capabilities of AI models and, in particular large language models, there is a strong focus on verifying their behavior is reliable, trustworthy and robust. However, once-off verification of AI models is not enough, as models are often open sourced and can be modified and deployed in different application context where the original behavior guarantees might not hold. In this work, we investigate the issue of when AI certification might cease to be valid in the AI life cycle. We propose leveraging use case similarity to assess whether original behavior claims hold and use knowledge graphs as a representation of a use case which allows modeling and easy modification of a broad use case context. We showcase on a real-world examples how use cases represented as knowledge graphs can be compared to understand whether original AI verifications hold.

Submission Number: 14