Designing Human-AI systems that work: turning everyday use into training data

Description

In today’s rapidly evolving technological landscape, Intelligent AI-powered features have woven themselves into every aspect of our digital experiences.From entertainment platforms that personalize what we watch or listen to, to feeds across social media, from voice assistants to intelligent copilots and many more. These systems often look impressive in demos, yet plateau in real use. The limitation is rarely model size alone. More often, it is the absence of well-designed human-AI feedback loops that allow systems to learn continuously from actual usage. Human-AI feedback loops are the mechanisms through which AI systems evolve based on real user behavior and outcomes. When designed well, they transform routine interactions into reliable learning signals. The most effective loops combine two complementary inputs: explicit feedback, where users intentionally express preferences or corrections, and implicit feedback, where user behavior reveals acceptance, effort, or friction without added burden. This talk presents a practical, human-centered playbook for designing feedback loops into AI systems. It outlines how to identify high-signal explicit feedback at natural decision points, how to interpret implicit behavioral signals such as acceptance, rejection, and downstream outcomes, and how to combine both to create attributable, high-fidelity learning signals. The framework also addresses when and where feedback should be collected to minimize disruption and enhance ux.