Work

How good are people at discerning AI-produced accents of English?

Year: 2026
Type: preprint
Abstract: Previous research in the evolutionary and linguistic sciences suggests that accent functions as a marker of social identity. Yet these signals are exploitable through effective signalling. We previously showed that native speakers of local accents in the UK and Ireland are effective at discerning fakery overall and fakery specific to their native accent. The advent of generative AI (genAI) tools, however, may make accent mimicry easier in the digital space, and reports suggest that genAI fakery may be an effective tool for exploiting listeners through scams. In this study, we build upon our work to investigate the degree to which native and non-native speakers of accents spoken in the UK and Ireland can detect genAI-produced accent mimicry. Following our previous experimental design, we used ElevenLabs – a genAI tool specializing in voice creation – to generate fake stimuli for seven accents across the UK and Ireland. Native listeners validated the quality of these stimuli. We then ran an online experiment through Qualtrics (N = 2000) to determine whether listeners are effective at discerning genAI-produced mimicry both overall and by accent-subgroup. We present an interim findings at the Cambridge Language Sciences 2026 meeting.
Language: English
Field: Psychology
Open Access status: green
OpenAlex ID: W7235981296
Cites:
Cited by:
Related to: 10