Upol Ehsan on Human-Centered Explainable AI and Social Transparency

Explainability

Upol Ehsan emphasizes the importance of explainable AI (XAI) in today's decision-making processes, which are increasingly powered by AI systems. He explains that XAI involves techniques and strategies that help stakeholders understand why an AI system made a particular decision, highlighting the human-centered aspect of this technology 1. Ehsan differentiates between transparency, interpretability, and post hoc explanations within XAI, noting that transparency involves making models clear, while interpretability requires expertise to scrutinize algorithms 2.

Explainable AI refers to the techniques, the strategies, the philosophies that can help us as stakeholders with an AI system.

--- Upol Ehsan

He also mentions the challenge of explaining deep learning models, which are often complex and not inherently interpretable 1.

   

Pitfalls

Ehsan discusses common pitfalls in explainable AI, such as the overtrust in numerical explanations, even when users do not fully understand them 3. He highlights that these pitfalls can occur without malicious intent, similar to natural pitfalls in the real world, and stresses the importance of recognizing and mitigating them 4.

What happens when harmful effects emerge when there is no bad intention behind it?

--- Upol Ehsan

Ehsan's work aims to articulate these pitfalls and develop strategies to address them, contributing to a more robust understanding of XAI 3.