Voice-AI assistants offer innovative ways for people to interact with technology, such as delivering search results through human-sounding voices. Unlike printed text, however, voices are associated with particular characteristics, such as accents, which have the potential to influence perceived credibility. Voice-AI assistants are also a relatively new phenomenon, and while people who use them frequently may be inclined to trust them, other people may not be. The current study investigated how voice accent and frequency-of-use affected users’ credibility assessments of search results delivered by voice-AI assistants. Participants, who were native speakers of American English and self-classified themselves according to how frequently they used voice-AI assistants, listened to statements produced by neural text-to-speech in either an American English or British English accent. They then rated the credibility of both the information content and the voice itself, along several dimensions. Results showed that in multiple conditions, participants perceived information delivered by British-accented voices as more credible than that delivered by American English-accented voices. Furthermore, frequency-of-use exerted a significant effect on perceived trustworthiness of a voice. These findings have implications for the ethical design of voice-AI systems, and for human-computer interaction more generally. Results showed that in multiple conditions, participants perceived information delivered by British- accented voices as more credible than that delivered by American English- accented voices. Furthermore, frequency-of-use exerted a significant effect on perceived trustworthiness of a voice. These findings have implications for the ethical design of voice-AI systems, and for human-computer interaction more generally. KEYWORDS voice-AI, speech perception, accents and dialects, trustworthy AI, credibility 1 Introduction In the early 2000s, internet technology underwent a revolution, as search queries began providing results whose quality and relevance were unprecedented. This association was clearly reflected in language attitudes in the 1980s, when Stewart et al. (1985) reported that “received pronunciation” (RP) accents of British English received the highest favorability ratings in the English-speaking world, including among speakers in the U.S. More recently, however, Bayard et al. (2001) reported that American accents were perceived as most favorable, and concluded that “the American accent seems well on the way to equaling or even replacing RP as the prestige – or at least preferred – variety” (p. 22). However, this conclusion is far from settled. In a follow-up study that adopted different methods from those used by Bayard et  al. (2001), Garrett et  al. (2005) reported that American participants overwhelmingly associate British English with positively “cultured” adjectives such as intelligent, refined, and well-spoken, although these results were tempered by associations indicating negative affect, such as snobbish. More recently, Wolfram and Schilling (2016) argued that North Americans continue to place value on British accents, speculating that this may be due to “a lingering colonial effect” (p. 34). Meanwhile, van den Doel (2006) has shown that U.S.-based listeners rated pronunciation errors (such as wea[d]er for weather) very differently when they occurred in a British accent, compared to in an American accent. Following the design of Gaiser and Utz (2023), participants listened to statements that were either of high accuracy (e.g., although e-cigarettes are less harmful than traditional cigarettes, they still contain carcinogenic substances) or low accuracy (e.g., the risk associated with smoking a hookah is significantly reduced compared to smoking cigarettes). The statements were produced by neural text-to-speech in either an American English or British English accent. Furthermore, we also anticipate that participants will perceive smaller differences in credibility for high- versus low-accuracy statements produced by British accents, compared to American accents. This is based upon the findings of Gaiser and Utz (2023), where an overall effect for voice-AI versus text statements was accompanied by a significant interaction with accuracy. The response options were: Never, Rarely, Once a month, Weekly, Daily. 3 Results Participants’ responses to the usage question were coded binarily as either Frequent device usage (daily, weekly, monthly; n = 82) or Infrequent device usage (never, rarely; n = 117). The responses to each of the eight rating questions were analyzed with separate mixed effects linear regression models using the lmer function in the lme4 R package (Bates et al., 2015). Each model included fixed effects of Accuracy Level (higher vs. lower), Speaker Accent (British vs. American), and Listener Device Usage Frequency (Infrequent vs. Frequent) and as well as all possible two- and three-way interactions. Low- versus high-accuracy statements received significantly different ratings for accuracy when they were delivered in a British English accent (coef. = 8.1, SE = 1.5, t = 5.4, p < 0.001), but the difference in ratings was even larger when they were delivered in an American English accent (coef. = 12.0, SE = 1.5, t = 7.9, p < 0.001). There was also a significant interaction between Voice Accent and Listener Usage Frequency (coef. = −1.1, SE = 0.5, t = −2.3, p < 0.05). This is depicted in Figure 2. Infrequent users of voice-AI rated British English statements as more accurate than American English statements (coef. It seems unlikely that a native speaker of American English would fail to detect the presence of a British accent, although it is certainly plausible that participants had different levels of conscious awareness about it, and that these differences could affect their credibility ratings. This issue could be explored in future work. Meanwhile, participants