An Efficient Approach for Enhancing GenAI Trustworthiness

As Generative AI (GenAI) and Large Language Models (LLMs) become increasingly integral to a wide array of applications, ensuring their trustworthiness has emerged as a critical concern. This paper proposes a new approach to enhance GenAI models' trustworthiness by returning for any given output its “reference list,” i.e., the training instances closest to it. Specifically, we show how the proposed approach can mitigate two major challenges of models' truthfulness: the tendency for hallucination and misinformation. Thus, by creating an equivalence to the human referencing practice, the proposed approach provides a step toward AI that humans can trust. The approach is generic, does not require access to the model, and uses dimensionality reduction and clustering of the training data to be efficient in sub-linear time.