This research focuses on the novel question of whether filled pauses and other disfluencies might be usefully inserted into AI-generated speech in educational contexts and thereby serve a facilitative function for learning. This work began by looking closely at the ethics of AI and took the current consensus of this area – that AI ethics revolve around five principles: beneficence, non-maleficence, autonomy, justice, and explicability – in order to develop and propose a useful tool for specifically evaluating the use of disfluencies in AI-generated speech for learning contexts. This tool is called the Disfluency Instrumentality Audit (DIA) and can be used by system designers and other researchers to evaluate whether the systems they are designing meet basic ethical standards.

In further work related to this research, two pilot experiments were carried out. One looked at whether deep learning systems (e.g., OpenAI’s chatGPT) actually emulate some of the facilitative effects often observed in previous research on disfluencies. A second pilot experiment was designed to look at whether disfluencies may facilitate a listener’s comprehension of texts. Overall, the work on this project suggests that the use of disfluencies in AI-generated speech for learning may be both productive and ethical.