About
Large language models (LLMs) like GPT-3 are revolutionizing artificial
intelligence, leading to breakthroughs in question answering, natural language
understanding, and machine translation. Recent work in a
variety of social science disciplines, including psychology, economics, and political science
has demonstrated remarkable similarity between the behavior
of LLMs and human decision makers. At the same time,
AI researchers and engineers struggle to understand these
systems, leading to practical challenges and ethical questions
about fair and safe deployment.
This workshop aims to bring together researchers to
discuss work on using psychological methods to understand
LLMs and LLMs as tools for understanding humans. Along
these lines, we have invited leading researchers from
cognitive science, psychology, and machine learning to
present their work on topics that include: When and why do
LLMs exhibit biased behavior? How do these compare to
human biases? What sorts of psychological tasks do LLMs
struggle with? Can we use psychological theory to structure
this search? And how does the knowledge encoded in LLMs
differ from human knowledge?
We expect these topics to be relevant to cognitive,
developmental, and social psychologists, behavioral
economists, sociologists, linguists, philosophers, computer
scientists, and AI safety researchers. We also believe that
the theme of this workshop is especially relevant to this
year's focus on "Cognition in Context". LLMs are rapidly
being deployed in many industrial settings and products,
and psychological methods may be key to understanding
these models and improving their performance. Cognitive science may thus
have a crucial role to play in the development of safe, robust
artificial intelligence and systems, providing crucial guidance
to government regulators and policymakers.
Structure
This workshop was a hybrid event with both in-person and virtual components.
We first held a half-day in-person workshop during CogSci 2023 in Sydney.
All in-person talks were recorded and released below along with pre-recorded talks from the virtual
speakers.
We then held a virtual component, featuring an
interactive town hall on Zoom discussing challenges opportunities for using LLMs for cognitive science.
You can view a recording of this town hall below.