From Human-Like Reasoning to Memorization Biases: When Should We Align AI Models?

Speaker: Afra Mashhadi (University of Washington Bothell)
Host: Gianluca Demartini

Abstract: As AI systems increasingly approximate aspects of human reasoning, a central tension emerges: When should we push models toward more human-like inference—and when does such alignment risk amplifying hidden biases or memorization artifacts? In this talk, I examine this question by unifying two strands of my recent work: the development of Theory-of-Mind (ToM)–centric approaches to LLM reasoning, and empirical analyses of memorization and representational biases in large-scale language models. I will present new findings from our study on social reasoning using Reddit data, illustrating where LLMs succeed—and fail—at inferring intent, social dynamics, and situational context. I then turn to our research using global scholarly co-authorship networks, where we show how memorization patterns are unevenly distributed across regions and disciplines, revealing deeper structural biases in what and whom LLMs remember. Together, these results highlight fundamental trade-offs in building more “human-like” AI systems and raise the question of when alignment is beneficial, when it is risky. 

Bio: Afra Mashhadi is an Associate Professor in the Computing and Software Systems Division at the University of Washington Bothell. Her research focuses on trustworthy and responsible artificial intelligence, with an emphasis on human-centric reasoning, data-centric evaluation, and the analysis of memorization and representational biases in large language models. Her work seeks to understand how AI systems interpret human intent, how they encode and reproduce global knowledge structures, and how these processes give rise to disparities in accuracy, representation, and downstream decision-making. Professor Mashhadi’s scholarship spans artificial intelligence, computational social science, and large-scale data analysis, and has been published in leading venues including WebSci, AIES, CIKM, CSCW, and FAccT. She is the recipient of support from the National Science Foundation and has contributed to research and industry innovation through previous appointments at Bell Laboratories and University College London. She holds a Ph.D. in Computer Science and has extensive experience leading interdisciplinary projects focused on the development of transparent, robust, and human-centered AI systems