Insights from Penn’s AI Experts

The information ecosystem—comprising social media, print and digital publishers, national and local television, streaming services, and radio—is vast in scale and heterogeneous in nature. As such, it is extremely difficult to study even descriptively: simply answering the question "what did the media report today?" requires synthesizing data from potentially tens of thousands of sources and many different formats. Addressing more complex questions about the prevalence of false, biased, or otherwise misleading information as well as its effects on public understanding and opinion is even more challenging. Fortunately, LLMs are now capable of ingesting and labeling natural language at human level quality and at scales that plausibly allow us to characterize the entire information ecosystem in close to real time. This change, which has happened in only the last year or so, will transform the study of media, technology, and democracy, with implications for journalism, social media platforms, and policy makers. 

I am studying how AI complements and substitutes labor and how AI and robotics help workers to become more productive. I also examine how firms use AI as a strategic asset to innovate new products and processes. In this stream, I also examine how AI affects competition in industries such as in drug-discovery and app development.

Humans have been thinking about artificial intelligence for thousands of years, and trying to create robots and sentient AIs for almost as long as that. The historical record across multiple continents is filled with legends, myths, and histories of attempts to create artificial life or artificial thought. Many of the examples from the western intellectual tradition, going back as far as Homer, appear as tools for surveillance, policing, and other forms of social control to serve the interests of the powerful with perfect obedience. How have these desires and fantasies, which emerge in the context of widespread enslavement and political absolutism, continued to influence our ideas about what—and who—this kind of technology is for?

Synthetic Design and Retrosynthetic Analysis will greatly benefit from AI. The same is true for interactive chemical structure elucidation based on NMR, IR, UV-VIS and mass spectra. I also see potential applications in the biosynthetic classification of natural products.

GenAI's responses are random and can lead to fairness, privacy and other societal issues. These topics are within my research areas of data science.

This can be through better tools for neural and behavioral data analysis but, more interestingly, as models of the brain itself that can help us develop new theories for how we learn and represent information.

Due to the proliferation of media stories about vulnerabilities and potential harm of AI, bridging the trust gap between developers of AI technology and skeptical users is critical to realize the promise of AI.  Penn Engineering's recent initiative in Trustworthy AI—ASSET (AI-Enabled Systems: Safe, Explainable, and Trustworthy) Center—is focused on developing the foundations necessary to bridge this gap.

We have developed new hardware where the individual components (adjustable nonlinear resistors) adjust according to local rules, allowing the system to learn supervised tasks as a collective physical response to inputs. Because learning is distributed among the components, and physics performs the inference in our system, our approach is far more efficient in energy and time.

Searches for new physics at colliders always make some assumptions about the nature of the new physics, typically in the form of specific models or classes of models that would give rise to a hypothesized new particle. However, if these models are not representative of the true new physics in nature, then they are incapable of discovery. Anomaly detection has the potential to allow for searches for new physics without imposing these priors on what the new physics might look like. Using these unsupervised learning techniques most effectively will require development on both the algorithm side as well as on their application to searches, most notably in the form of uncertainty quantification.

In the Mathijssen lab, we use physics-informed neural networks (PINNs) to design active materials and to study the contamination dynamics of bacteria. This technology can help us with image processing and data analysis, but only if used appropriately. 

We combined artificial intelligence (AI) with stream-of-consciousness to make the latent construct of personal narrative explicit. Adult participants (n = 26) contributed 50 stream-of-consciousness thoughts, which along with demographic details and our prompts, were processed by ChatGPT-4 to create a personal narrative. Participants evaluated these AI-generated narratives for accuracy, surprise, and insightfulness, reporting high accuracy, surprise, and increased self-insight. Twenty-five of the 26 participants rated the narratives as ‘Completely Accurate’ or ‘Mostly Accurate’, 19 rated the narratives as ‘Very Surprising’ or ‘Somewhat Surprising’, and 19 indicated that they learned something new about themselves.

All important questions are ultimately about causal effects. AI can and should thus integrate insights from causal inference.