AI Fellows discuss AI opportunity, hazard at Arts & Sciences Showcase

Artificial intelligence is reshaping science and scholarship in ways researchers are just beginning to grasp. That was the unmistakable message of the AI Arts & Sciences Showcase, held on April 15 as part of AI Month at Penn, where eight graduate students and postdoctoral researchers delivered rapid-fire talks ranging from the subatomic to the societal.

Several talks examining AI in the context of human learning and culture noted reasons for caution. Margaret Kandel (Linguistics) asked whether children can learn language from AI chatbots and found that while they produce longer remarks than humans, their language tends to be linguistically impoverished, less varied in vocabulary and structure, and less grounded in shared context. In the same vein, Fabian Baumann (Biology) discussed how the increasing delegation of content creation to a handful of generative AI models leads to more and more of the text and images we encounter sharing remarkably similar features. Over time, using these AI generated contents to train new models further decreases the diversity and quality of model outputs, leading to “model collapse” for the AI algorithms and impoverishment of cultural products for humanity. And even the factual contents of AI outputs require scrutiny: Marvin Maechler (Psychology) argued that confirmation bias may be an inescapable feature of any system that maximizes accuracy while learning under noise.

Other researchers showcased AI as a powerful instrument for discovery. Max Cohen (Physics & Astronomy) described a neural anomaly-detector used for screening particle collision data at 40 million events per second to hunt for new physics in real time, while Ben du Pont (Physics and Astronomy) used generative models to design novel molecules by steering samplers into uncharted chemical space. Carolina Torreblanca (Political Science) deployed machine learning to comb through over 91,000 articles and track the “credibility revolution” in her field, and Elena Liang (Data Driven Discovery) predicted illegal dumping hotspots across Philadelphia, using data to inform concrete community interventions. Finally, Shreya Arya (Mathematics) introduced “MathDuels”, a framework in which AI models generate problems for each other, as a rigorous new approach to evaluating mathematical reasoning.

Taken together, the talks made clear that some of the most important opportunities and pressing questions presented by AI will require the full breadth of perspectives across the sciences and humanities. It is equally important to learn the mathematical underpinnings that steer the algorithms as it is to understand how they interact with and impact the people using them and the broader cultural landscape. Knowing how to harness these new tools efficiently and ethically can propel scientific discovery, and our AI fellows are helping the whole Penn community to chart that path.

The AI Arts & Sciences Showcase was co-hosted by Penn Arts & Sciences’ Data Driven Discovery Initiative, MindCORE, Price Lab for Digital Humanities, the Center for Soft & Living Matter, and PDRI/DevLab.