UC Merced Magazine | Volume XX, Issue VI

It’s not exclusive to humans, though. Some primates have this ability, and researchers only recently have been able to ask whether AI might also exhibit the trait. “Do large language models have theory of mind?” Kello asked, referring to the neural networks that underlie soaringly popular chatbots like ChatGPT. “Well, they exhibit many of the behaviors taken as evidence for it.” Studies have subjected LLMs to a benchmark test that measures theory of mind. By some measures, the AIs passed many of the tests “with ying colors,” Kello said. e implications of studies showing similarities between humans and AI are leading researchers down numerous avenues of inquiry. Kello and Rachel Ryskin, another UC Merced cognitive science professor, are working on a study that examines the degree to which LLMs can adjust their responses based on an understanding of human memory and attention — one of the many earmarks of theory of mind. Results show as they are trained on massive collections of data, LLMs learn about human limits of memory and attention and can adjust responses accordingly, even though LLMs hardly share those limitations. For Kello, these results refer to a vitally important issue called alignment, which is a desire for AI to align with human intentions, limits and values. It has been known for years that AI algorithms learn human biases through their training data. Extraordinary e orts have been made to root out these biases in chatbots and other AI technology. Some researchers, however, believe this problem may become existential as AI’s capabilities evolve. “We’ve come so far so fast, it doesn’t take a lot to imagine wiring up a robot with some reinforcement learning so they can function in the world but behave in a way we can’t directly control,” Kello said. “ is is where the existential danger comes in.

“Part of my value system is grounded in my experiences over time, of being in a culture. We call that conceptual alignment. We need to be sure AIs are grounded in the same conceptual framework we are.”

Professor Chris Kello

Professor David Noelle

We’ve come so far so fast, it doesn’t take a lot to imagine wiring up a robot with some reinforcement learning so they can function in the world... — Professor Chris Kello

(Continued on page 22)

21

UC MERCED MAGAZINE // ucmerced.edu