Can Chatbots Help Us Better Understand People?

Graphic with dark background with white, red, and pink electrical currents passing in the background. "CAOE Seminar Series: Perils and Promises of Artificial Intelligence for the Homeland Security Enterprise" is written across the graphic.

Thursday, November 21, 2024

12:00 PM (EST) | 10:00 AM (MST) | 9:00 AM (PT)

About the Seminar: Can Chatbots Help Us Better Understand People?

Large Language Models powering modern chatbots have demonstrated impressive capabilities in conversational AI and task-solving. However, researchers also worry about the risks of harm, such as propagating historical biases embedded in the data used to train chatbots. This talk offers a novel perspective: rather than treating bias as a risk to be minimized or eliminated, I show how to leverage it to gain insights into people. First, I demonstrate that it is possible to skew a chatbot’s ideological perspective on a given topic with minimal training data. Surprisingly, altering its perspective on one topic makes the chatbot sound like a partisan on other topics as well. Next, I show that this manipulation strategy can be used to align chatbots to human populations, effectively creating “digital replicas” that mirror their language, style, and biases. These digital replicas can be “surveyed” by prompting population-aligned chatbots to reveal the perspectives, beliefs and even psychological states, of diverse populations. While this work opens new possibilities for studying human behavior and social dynamics at scale, it also raises critical questions about fidelity, validation, and the ethical development of AI systems capable of emulating human perspectives.

Dr. Kristina Lerman

Senior Principal Scientist at the Information Sciences Institute and Research Professor in the USC Viterbi School of Engineering’s Computer Science Department.

The Center for Accelerating Operational Efficiency (CAOE) invites you to our seminar series Perils and Promises of Artificial Intelligence for the Homeland Security Enterprise. This seminar in this series will feature speaker Dr. Kristina Lerman, Principal Scientist at the Information Sciences Institute,a unit of the University of Southern California (USC) and a Research Professor in the USC Viterbi School of Engineering’s Computer Science Department. An expert in complex multi-agent systems, Dr. Lerman’s research on social data and other topics has been funded by the National Science Foundation (NSF), Defense Advanced Research Projects Agency (DARPA), Airforce Office of Scientific Research (AFOSR) and the Army Research Office (ARO).

Her current work revolves around deciphering the structure and dynamics of social media and crowdsourcing platforms, such as Twitter, Reddit, and Stack Exchange among others. Among her goals: understand the role of networks and plaform’s content curation algorithms in shaping collective behavior, discover the structure of user-generated communities, predict emerging trends and group behavior, and identify the role of cognitive constraints in online interactions. Projects explore network-based and machine learning-based approaches to harvest concept hierarchies from social metadata and automate semantic annotation on the social web.

Dr. Lerman’s research includes statistical text analysis, semantic modeling of data, and mathematical modeling of multi-agents systems, as well as social networks and social computing. She has published a variety of journal articles, book chapters, and refereed conference and workshop proceedings, and has written numerous workshop papers and technical reports. She also holds a patent involving document-based data extraction.

Dr. Lerman teaches a USC Computer Science Department course on social media analytics. She briefly worked for a tech startup prior to joining ISI. Dr. Lerman was awarded her Bachelor of Arts in Physics by Princeton University, and her Ph.D. in physics by the University of California at Santa Barbara. 

About the CAOE Seminar Series

Artificial Intelligence (AI) has taken the world by storm with applications in self-driving cars, medical diagnostics, biometric analysis, and more. However, the use of AI is not without its own drawbacks. As such, it is critical to understand the current state of the art in AI, where the field is going, and what can (or cannot) currently be accomplished with Artificial Intelligence. In this seminar series, we invite AI luminaries to discuss current advances in Artificial Intelligence and offer insights into the perils and promises of such technology for the Homeland Security Enterprise.

Event Location

Attend: via Zoom webinar