How does working memory predict errors in human-AI interaction?

Interlingual Respeaking (IR) is a new technique that enables real-time subtitling in a different language. This cognitivelydemanding technique involves collaboration between a language professional and automatic speech recognition software(ASR), creating a human-AI interaction (HAII) environment. Integrating technological tools with an individual’s internalcognitive resources establishes an extended cognitive system. However, different types of errors are observed in termsof output accuracy. Our ESRC-funded research found that working memory (WM) (backward span) has a negativerelationship with omissions, where content is dropped out (e.g., to save time). Nevertheless, additions, where the humanadds content (e.g., to clarify meaning) and correctness, where form-related issues arise (such as grammar mistakes), hadan inverse relationship with the N-back Task (the simultaneous maintenance, updating, and processing of WM). Thesefindings suggest that the IR errors involve diverse types of WM resources.