Responsible AI in the World of Work
Speaker: Fabio Casati
Abstract: When can we say that an AI model or pipeline is “reliable” – literally meaning that we can rely on it? How do we evaluate and choose the most reliable one? How do we take go / no go decisions for our business? These are only some of the questions that AI producers and consumers ask (and ask themselves) every day. Gen AI is now mainstream, but most process owners, even in large enterprises evaluate systems either anecdotally or using metrics and methods with unclear correlation with AI quality. Based on dozens of interviews and uncountable interactions with developers, managers, customers, and institutions, in this talk I summarize what I perceive as the key issues in AI quality evaluation and decision making. I will outline the different personas, their perception of who is responsible and accountable for decisions, and the relevant quality and decision factors. I will also discuss risk, and how risk is assessed by different actors at different stages of the production pipeline. I will stress why commonly used practices are not just insufficient, but outright misleading and how confusion in terminology can and does lead to wrong decisions. Finally, I will comment on frequent pitfalls in experiment design and reporting, again resulting in leading managers to take the wrong decisions, and how to correct them.
Short Bio: Fabio Casati is a AI Architect at Servicenow as well as technical lead for AI Trust and Governance. Fabio focuses on designing, architecting and deploying AI-powered workflows for enterprise customers. On the research side, he is working on AI applied to workflows and on quality in AI. Previously he was Professor at the University of Trento. In that role, he started research lines on crowdsourcing and hybrid human-machine computations, focusing on applications that have direct positive impact on society through tangible artefacts adopted by the community. Prior to that, he was technical lead for the research program on business process intelligence in Hewlett-Packard USA, where he contributed to several HP commercial products in the area of web services and business process management. He co-authored a best-selling book on Web services and is author of over 250 peer-reviewed papers and dozens of patents.