Examining Scientific Advances via the Socio-Technical System of AI

From AI risk management and physics-informed models to computational biology and scholarly publishing, five leading voices joined the Northwestern Network for Collaborative Intelligence (NNCI) this fall to examine what AI can and cannot do, and how people should use it responsibly. Coming from wide-ranging vantage points, the distinguished speakers converged on a shared theme: AI’s biggest challenges are human as well as technical.

More than 400 researchers, students, and practitioners attended the five Distinguished Speaker Series events, through which NNCI aims to spark interdisciplinary and cross-functional dialogue and illuminates emerging ideas around responsible and high-impact research and education in data science and AI.

Perspectives on AI Risk

On Sept. 28 at NNCI’s Fall Distinguished Speaker Series kickoff event, William Regli discussed AI as a socio-technical system. He outlined an approach to AI risk that recalibrates the engineering problems—like fixing the algorithm, cutting hallucinations, and improving benchmarks—with AI’s thorniest non-technical problems, namely its social, behavioral, and cognitive harms.

With decade-long experience spanning multiple public service roles in AI and national security, Regli is a professor of computer science and founder of the Applied Research Laboratory for Intelligence & Security (ARLIS) at University of Maryland, College Park.

Discussing threat amplification and externalities such as IP scraping, energy use, and pollution as significant near-term concerns, Regli argued that AI risk should be managed like cybersecurity or the weather. In that framework, users sort risks by the known and unknown, then apply a prevent-detect-respond-and recover response.

Learning Cellular Organization and Function Across Scales

NNCI’s Sept. 29 Summer Series event was co-organized by Professor Feng Yue and the Northwestern University Feinberg School of Medicine. Guest speaker Jian Ma addressed a central challenge in modern biology: integrating heterogeneous measurements across molecular, cellular, and tissue scales to build a predictive understanding of cell diversity and tissue function. Arguing that today’s scale AI models trained on large genomics datasets do not produce new biological insights, Ma cautioned that models may simply be capturing the correlations between the data modalities versus the underlying mechanisms for causal connections.

Ma is the Ray and Stephanie Lane Professor of Computational Biology and founder of the Center for AI-Driven Biomedical Research at Carnegie Mellon University. He presented recent research in his lab, including models that link three-dimensional genome folding to gene regulation in Alzheimer's disease, and a multi-scale attention model called Steamboat that captures how cells influence one another across different spatial ranges within tissue.

Neural Operators: AI that Understands the Physical World is Essential for Scientific Discovery

On Oct. 21, NNCI and Northwestern Engineering's Department of Computer Science co-sponsored a NSF-Simons AI Institute for the Sky colloquium featuring Anima Anandkumar.

Anandkumar, the Bren Professor of Computing and Mathematical Sciences at the California Institute of Technology, discussed the limits of purely data-driven AI. She pioneered neural operators, a class of AI models that learn multiscale phenomena common in nature, such as fluid dynamics, material behavior, and wave propagation. This class of models approaches scientific problems as mappings between function spaces rather than fixed grids, incorporating physics and mathematical understanding to accelerate scientific discovery.

Using neural operators, Anandkumar trained the first high-resolution AI weather model, which runs tens of thousands of times faster than traditional forecasting. In 2023, the FourCastNet model correctly predicted Hurricane Lee's Nova Scotia landfall days before conventional models. Anandkumar’s team adapted the model’s core neural architecture, to simulate black hole mergers and nuclear fusion reactors.

A Trinity Platform for Scholarly Publications: A Harvard Experiment

On Oct. 22, NNCI welcomed Xiao-Li Meng, the Whipple V. N. Jones Professor of Statistics at Harvard University and the founding editor-in-chief of the Harvard Data Science Review (HDSR). Meng explored scholarly publication in the age of AI, emphasizing HDSR’s three pillars of content, community, and a shared culture of intellectual integrity, rigorous pursuit, and critical evaluation.

Discussing credit and authorship, Meng argued that AI contributions should be explicitly acknowledged, not hidden, while accountability remains with human authors. He also called for the better credit and preservation of peer review reports, which Meng described as an untapped, richly human-labeled dataset on decision-making.

Meng articulated his goal for HDSR as “everything data science, for everyone.” Reflecting on how conversations with a philosopher and a library/information scientist expanded his own view of data science, Meng proposed that data is never neutral or raw. He explained that all data is human processed.

An Evening with Dennis Chookaszian

Dennis Chookaszian (’65), a Life Trustee at Northwestern and a member of the Global Advisory Board at Kellogg and the McCormick Advisory Council, joined NNCI on Oct. 28 for a lecture and moderated discussion with Professor V.S. Subrahmanian. Drawing on his experience in leadership roles across industry, finance, and entrepreneurship, Chookaszian shared advice on AI use and implementation. He emphasized that academic institutions need to teach AI etiquette and then allow students to build their own AI use workflows, including how to prompt, verify, and edit for form and function.

The former Chairman and CEO of CNA Insurance Companies, Chookaszian served in various roles in his 26 years with the company, including president and chief financial officer. He is currently a director of two public companies, twenty private corporations, and six nonprofit organizations. Chookaszian also teaches courses in corporate governance and AI.