Voice assistants that over-explain win more goodwill than those that barely adjust

Voice-based assistants have quietly moved out of the kitchen and into roles that once belonged to people: tutors walking students through problems, advisors offering medical guidance, agents fielding customer service calls. As these systems take on relational work, users stop treating them as search boxes and start judging them the way they judge human conversational partners. A new experiment suggests that one of the most consequential judgments hinges on how the machine speaks, not just what it says. When a voice assistant talked too much, too slowly, and too simply, listeners rated it more favorably than when it rushed, used complex vocabulary, and offered minimal clarification.

The study, published in the journal Communication and Change, was conducted by Varun Rijhwani, Matthew Craig, and Chad Edwards. It drew on two theoretical frameworks: Communication Accommodation Theory, which describes how communicators adjust their behavior to manage comprehension, social distance, and relational alignment, and the Stereotype Content Model, which holds that people evaluate social targets primarily along two dimensions, warmth and competence. The researchers wanted to know how these evaluative lenses apply when the communicator is a synthesized voice rather than a human being, and whether the answer changes depending on the role the assistant is playing.

Their experiment recruited 284 adults from the United States through the Prolific platform in November 2025. Participants, who ranged in age from 19 to 82 with an average age of about 45, each received one dollar for completing a survey that took a median of five and a half minutes. Most identified as White, and the sample was roughly balanced between men and women with a small nonbinary group. After providing informed consent under a human subjects review board approval, each participant was randomly assigned to one of six conditions in a two-by-three between-subjects design that crossed accommodation style, overaccommodative versus underaccommodative, with relational context, tutoring, medical advice, or customer service.

The stimuli were six scripted audio vignettes. Each began with a brief introduction asking participants to imagine interacting with a voice-based agent similar to Amazon Alexa or Apple Siri in one of the three roles, performing a routine task such as receiving academic help, medical guidance, or service assistance. In the overaccommodative condition, the assistant spoke more slowly, used simplified language, repeated information, and provided excessive clarification. In the underaccommodative condition, it spoke more rapidly, used more complex vocabulary, and offered minimal clarification. The scripts heightened the mismatch by having the assistant maintain its style even when simplification was unnecessary or when the listener was clearly struggling. Crucially, the substantive informational content was held constant across all conditions, and the audio clips were matched as closely as possible in duration, differing by only a few seconds, so that only vocal delivery and role varied.

The manipulation was not perfectly clean, and the authors are candid about it. Participants rated how much the assistant over-compensated and under-compensated in adapting its communication. The undercompensation item worked as expected: those in the underaccommodative condition rated the assistant as under-compensating significantly more than those in the overaccommodative condition, with a moderate effect size. But the overcompensation item behaved oddly, with underaccommodative-condition participants actually rating the assistant as more over-compensating than those who heard the overaccommodative version, a small effect the authors attribute to a possible response assimilation effect in which any communicative mismatch leads listeners to judge the agent as poorly calibrated on both dimensions. The primary hypotheses, they note, rest on the experimental manipulation and outcome measures rather than on the manipulation check alone.

The main results were unambiguous. On a multivariate analysis combining warmth, competence, and perceived accommodation, accommodation style produced a large significant effect, while relational context and the interaction between style and context did not. Univariate tests showed that participants perceived the overaccommodative assistant as substantially warmer, averaging 5.09 on a seven-point scale compared with 3.22 for the underaccommodative version, a difference accounting for roughly a quarter of the variance. The overaccommodative assistant was also seen as more competent, 4.68 versus 4.13, and as more appropriately accommodative, 4.49 versus 3.25. In short, the assistant that overdid it was judged kinder, more capable, and better calibrated to the listener’s needs.

The pattern held for broader evaluations. Participants rated the overaccommodative assistant more positively as a communicator, describing the interaction as more satisfying, positive, and enjoyable, with means of 4.86 versus 3.03, and more positively as a speaker, judging it as more good-natured, warm, sincere, friendly, and trustworthy, with means of 5.21 versus 3.33. Both effects were large. Relational context, by contrast, barely mattered. It did not significantly change how accommodation style affected warmth, competence, appropriateness, or speaker evaluation. The one exception was communication quality, where a significant but modest interaction indicated that the strength of the accommodation effect varied somewhat across the tutoring, medical, and customer service settings.

The authors are careful about what this asymmetry does and does not mean. The design deliberately included no neutral or appropriately accommodative control condition, so the findings speak only to the relative consequences of two forms of nonaccommodation, not to whether overaccommodation beats normatively appropriate communication. Drawing on prior accommodation research, they argue that underaccommodation is consistently the most penalized form because it threatens comprehension and signals a failure to meet task-based and relational expectations, particularly in contexts involving expertise or support. Overaccommodation, meanwhile, may be read as well-intentioned or precautionary rather than patronizing. Because users do not attribute social intent to a machine, excessive explanation may be chalked up to design constraints, making overaccommodation a comparatively lower-risk strategy for voice agents.

The study also carries honest limitations. It used scripted audio vignettes rather than live, adaptive interactions, so it cannot say anything about accommodation as an iterative, turn-by-turn process that responds to user feedback, misunderstanding, or affect. The manipulation bundled speech rate, lexical complexity, repetition, and pacing together, so no single vocal feature can be credited with the effects. Participants experienced only a single exposure, leaving open how repeated interactions with persistently accommodating machines might reshape expectations over time, including the possibility, raised in the discussion, that users accustomed to unconditional machine accommodation may grow more sensitive to perceived underaccommodation from human partners. The authors also note that modern systems increasingly let users customize speech rate and verbosity, a factor the study did not examine.

What the experiment does establish is that the social costs of getting vocal tone wrong are asymmetric. A voice assistant that explains too much, repeats itself, and slows down pays little penalty in users’ eyes; one that speeds up, talks over their heads, and refuses to clarify pays a heavy one, and that penalty holds across tutoring, medical, and service roles alike. For designers of conversational AI, the practical implication is that when in doubt about how much to adjust, erring on the side of more support appears safer than erring on the side of less, at least as judged from a single scripted encounter. For researchers, the work extends accommodation theory into the auditory domain of human-machine interaction and flags communication quality as the one evaluative surface where relational context still leaves a fingerprint.

Subject of Research: How vocal accommodation style shapes social evaluations of voice-based assistants in relational roles

Article Title: Too much or too little? Vocal accommodation and social evaluation of voice-based assistants in relational roles

Article References: Rijhwani, V., Craig, M., & Edwards, C. (2026). Too much or too little? Vocal accommodation and social evaluation of voice-based assistants in relational roles. Communication and Change, 2(1), Article 15. https://doi.org/10.1007/s44382-026-00033-4

Image Credits: AI Generated

DOI: 10.1007/s44382-026-00033-4

Keywords: voice assistants, Communication Accommodation Theory, overaccommodation, underaccommodation, Stereotype Content Model, human-machine communication, warmth, competence, audio vignettes, social evaluation, Prolific, experiment