In order to release trustworthy AI products, we need to validate both how robust the products are when used with various inputs and how ethical their performance is. Although the importance of trustworthiness and ethics for AI is well recognized, seeking comprehensive validation measures and developing practical tools is still a common problem.
This session introduces how we select the validation measures and how we develop the validation tools for LINE's AI products, especially for large language models. For validation, we developed a stress test assessing fairness, toxicity, and unintended memorization for AI systems, employing adversarial machine learning techniques to efficiently generate test inputs.
The session welcomes all kinds of participants, including both those interested in AI ethics and beginners to the subject interested in learning more.
