Silence for the Machines
Shay Logan
2024/05/27 (mon) 10:00~12:00
Abstract:
Large language models (LLMs) like ChatGPT are enabling some surprisingly sophisticated chatbot-like applications of AI. Previous experiences with chatbots (e.g. Microsoft’s infamous experiments with Tay and Zo) have shown that they are often incapable of dealing with sensitive topics including race, gender, and ethnicity in non-offensive ways.
These problems were (and often still are) dealt with in a rather inelegant way: by blacklisting certain words or phrases. In joint work with Marcus Rossberg, Teresa Kouri, and Thomas Ferguson, I have been developing an alternative proposal based on conversational dynamics. At the heart of our approach is the idea of *intentional* silence—a speech act in which one actively avoids an utterance.
This talk will discuss our proposal in a generally nontechnical way. Along the way we will discuss ways intentional silence can impact our speech, how this might be modeled in a formal system, and how to implement such a model both in the context of conversational dynamics and in the context of LLM-generated content.
