Hussein Mozannar is a PhD student at MIT in Social & Engineering Systems and Statistics, advised by David Sontag. His research focuses on augmenting humans with AI to help them complete tasks more efficiently. Specifically, he focuses on building AI models that complement human behavior and designing interaction schemes to facilitate human-AI interaction.  Applications of his research include programming (GitHub Copilot) and healthcare (radiology and maternal health).

Attend this talk via ZOOM

AI systems, including large language models (LLMs), are augmenting the capabilities of humans in settings such as healthcare and programming. I first showcase preliminary evidence of the productivity gains of LLMs in programming tasks. To understand opportunities for model improvements, I developed a taxonomy to understand how programmers interact with a popular LLM extension, GitHub Copilot. This taxonomy reveals how much programming behavior has changed; particularly, time spent verifying LLM suggestions dominates other activities. I then show how we can leverage human feedback to improve the interaction. A key question is how does the human know when to rely on the AI or ignore its suggestions. I propose an onboarding procedure that allows users to have an accurate mental model of the AI for effective collaboration. However, in other settings where human resources are limited (healthcare), we might have to deploy AI selectively without human oversight. I show how to design AI systems that can predict on their own or defer the decision to humans when best to do so. Finally, I discuss how this line of work can build up to a vision of re-imagining human workflows with AI-enabled operating systems.