The dominant paradigm in machine learning evaluates systems on curated benchmarks, assumes large, labeled datasets, and treats deployment conditions as an implementation detail. This talk argues that paradigm breaks down in the settings where AI is most needed.

The failure is not incidental: it is structural. When costs are asymmetric, labels are scarce, and missed signals carry serious clinical consequences; the standard pipeline fails at the loss function, at the explanation, and at the evaluation metric simultaneously. The talk develops this argument through use cases in cardiac risk detection, neurovascular outcome prediction, agricultural pest detection, and energy anomaly identification, each chosen because it exposed a different layer of the same problem. That diagnosis makes the governance question inevitable: when these failures go unaddressed and AI supports consequential decisions, accountability cannot stop at model evaluation.

Scarcity, uncertainty, and asymmetric consequence are not exceptions to the problem. They define it.

Dr. Mariette Awad is a tenured Associate Professor in Electrical and Computer Engineering at the American University of Beirut and Director of AUB’s Artificial Intelligence, Data Science, and Computing Hub and the Human Machines Interaction Lab. She works at the intersection of efficient learning systems, explainable and responsible machine learning, and applied AI for real-world impact. She is the author of the widely cited and highly downloaded book Efficient Learning Machines and has been recognized among the Stanford–Elsevier top 2% most-cited scientists.

Her work advances deployable AI systems under constraints in data, compute, infrastructure, and governance. She has received the Abdul Hameed Shoman Award for Arab Researchers and the Intissar Ahdab Award for Women’s Excellence and Impact in Research. She has also contributed to AI policy and advisory work with UNESCO, UNICEF, UN-ESCWA, ITU, the International Science Council, and UNESCO’s Women for Ethical AI network.