Human AI at DDDI

The arrival of powerful tools of generative AI has sparked lively debate about how AI will impact scholarship and society at large. Penn’s DDDI aims to keep the human in Human AI front and center by fostering new research and interaction at all levels, from freshman to faculty. The School of Arts and Sciences stands out in the collaborations not only across related fields, but also between humanists, social scientists and natural scientists. DDDI aims to take our research and societal impact in Human AI to the next level.

AI is accelerating research into fundamental questions across the sciences, and opening up entirely new fields. AI related advances have also expanded the societal impact of Penn researchers. These trends are merely the beginning of a period of potential exponential growth in AI related research. Examples of recent research by Penn faculty serve to illustrate this potential:

AI and the mind – what does deep learning teach us about how we think? And how can we evaluate AI using insights from human psychology? Our psychologists use AI to understand how we reason, make social decisions, store memories and conceptualize the world around us (Anna Jenkins, Anna Schapiro, Sharon Thompson-Schill, Joe Kable). And, with colleagues in Biophysics, how we perceive the world around us, and more fundamentally what gives rise to human intelligence (David Brainard, Vijay Balasubramanian, Johannes Burge, Alan Stocker).

AI and language – Large Language Models have revolutionized both the modeling of language and how we as a society interact with AI. Penn social scientists and computer scientists use LLMs to study how we represent language in the brain, how messages in the media impact different social groups, and all kinds of ways that AI impacts human behavior – a brand new field (Sudeep Bhatia, Emily Falk and colleagues in Anneberg, Wharton, Lyle Ungar, Chris-Callison Burch, CIS).

AI and society – Sociologists and political scientists are using AI models to uncover intricate patterns from complex social data. Penn faculty sift through vast volumes of text from social media, administrative records, interviews, archival documents to identify shifts in public opinions, emerging cultural trends, political discourse, or nuances in social changes. Considerations of human ethics and equitable outcomes are a critical part of research and practice in these fields. (Xi Song, Dan Hopkins, Philip Tetlock, John Lapinski)

AI and justice – AI simultaneously holds promise and generates concerns for its role in the justice system. It may help find missing children and identifying perpetrators and at the same time exacerbate disparities or increase the risk of wrongful arrests. Penn researchers are using AI to reduce error rates in forensic toolmark analysis. They are studying misidentification rates in facial recognition algorithms when used on the typical law enforcement images (grainy CCTV and odd angles, rather than the clean images often used to train the systems). Machine learning forms a foundational component of common methods used for investigating and monitoring racial bias in policing due to innovations by Penn scholars. AI also is fundamental in establishing performance benchmark for police, prosecutors, and judges and can identify critical outliers, such as excessive use-of-force, race disparities in charging, and over-reliance on incarceration. (Greg Ridgeway, Maria Cuellar)

AI for networks – complex networks are ubiquitous: in our bodies, in nature, and in social media. Advancements in ML and AI have improved our ability to infer the structure and interactions across these networks and thus better understand how they function. How do blood vessel networks optimize the flow of nutrients through blood vessels in mammals, and what can we learn from rivers, leaves and other networks? How do social networks shape language, and how does language shape social networks? How do network structures and decision mechanisms interact to shape the transmission of disease in human populations, or the spread of gang violence through social media? AI has aided advances in these questions and in fact the broader study of animals in natural settings using continuous videography, an entirely new approach enabled by deep learning. (Eleni Katifori, Erol Akçay, Josh Plotkin, Marc Schmidt, Lyle Ungar, Desmond Upton Paxton)

AI for human cells – Biologists, physicists and electrical engineers at Penn are tackling key questions in the human genome, the structure of proteins, and the remarkable ways they exchange signals. The amount of available human genetics data will soon reach the scale of a billion individuals. Penn faculty use deep learning to uncover genetic factors responsible for human traits, predict DNA and RNA function, with applications to vaccines. New “multi-omics” approaches have revealed an underappreciated heterogeneity of cell types and states that impact how cells function. (Junhyong Kim, Sarah Tishkoff, Josh Plotkin, Brian Gregory, Eleni Katifori, Andrea Liu, Marc Miskin)

AI and climate – AI has enabled new approaches to forecasting complex climate dynamics with global consequences, and may become a powerful new tool accelerating innovations targeting the reduction of greenhouse gas emissions. Penn faculty combine rigorous physics-based research with AI to model climate, and study its impact on socio-economically diverse communities. At the same time, the computational demands of AI have raised questions about its carbon footprint, and created a new need to study its environmental impact. These opportunities and challenges require new policy and engineering to foster environmentally sustainable AI. (Michael Mann, Irina Marinov, Michael Weisberg, Deep Jariwala, Benjamin Lee, Penn Program on Regulation)

AI and the universe – The Penn astronomy group has a world leading program in applying AI to big data astronomy, in collaboration with theorists at the Center for Particle Cosmology. Recent discoveries include the largest comet ever found, advances in the cosmological puzzles of dark matter and dark energy, and ‘interpreting’ what deep learning tells us from galaxy images. Just in the past year, the group has been awarded the largest NSF and NASA grants awarded to any university group for machine learning applied to survey data from telescopes both on the ground and in space. They are currently proposing a national level AI Institute in Astronomy to the NSF (jointly with CMU and U Washington). (Gary Bernstein, Masao Sako, Bhuvnesh Jain, Robyn Sanderson, Dylan Rankin, Eric Wong, Jake Gardner)