Abstract The present paper uses a perception experiment to investigate which prosodic factors determine the perceived charisma and trustworthiness conveyed by phrase-final intonation contours. The following factors were manipulated using PSOLA resynthesis and combined into 16 stimuli: phrase-final pitch level (low vs. half-low), shape of fall (accelerated vs. decelerated), nuclear pitch-accent timing (H+L* vs. L+H*), and final lengthening (present vs. absent). Authentic statements by two U.S. politicians, a man and a woman, served as base stimuli. Forty native speakers of American-English rated the stimuli. Results showed that all factors influenced the listeners. The highest ratings were achieved by low and steep final falls, early H+L* pitch accent, and without final lengthening. Moreover, the rating levels for the female stimuli were generally higher but declined more under unfavorable phrase-final patterns. The results are in line with previous studies and are discussed with regard to practical implications and the notion of prosodic constructions.
Abstract People are vigilant toward the risk of misinformation and assess speakers’ trustworthiness to calibrate their trust accordingly. This assessment can target speakers’ epistemic commitments, conveyed through evidential claims. Speakers may take responsibility for the truth of a proposition by reporting first-hand evidence (“I saw that p ”) or defer responsibility by referring to second-hand evidence (“somebody told me that p ”). This pre-registered study experimentally investigates whether trust calibration is sensitive to the accuracy of these evidential claims. Our findings show that when a speaker shared accurate information p but provided an inaccurate evidential claim, the negative impact on perceived trustworthiness was greater for the strong perceptual evidential claim than the weak hearsay claim. These results provide new evidence that overcommitting by claiming direct visual evidence for information acquired second-hand is more reputationally damaging than appealing to hearsay despite having direct evidence. We discuss the implications for designing more trustworthy artificial agents.
Abstract This paper examines theoretical aspects of human-AI communication, arguing that current AI systems, or ‘pragmatics-zombies,’ can mimic human-like pragmatic competence without genuine understanding. While these systems are increasingly adept at natural language interaction, their lack of subjective experience, intentionality, and embodied cognition creates a fundamental trust deficit, particularly in high-stakes scenarios. We argue that forcing human-to-human pragmatic frameworks onto these non-human entities is suboptimal, leading to issues in efficiency, robustness, and transparency. To address this, we introduce the concept of ‘Abstract Pragmatics,’ a general framework that allows for multiple, distinct pragmatic systems. We propose that instead of pursuing flawless mimicry of human communication, research should focus on developing novel, alternative pragmatic systems specifically designed for human-AI interaction. This approach would prioritize clarity, verifiability, and trustworthiness, fostering a more effective and reliable collaboration between humans and artificial intelligence by acknowledging and adapting to the unique nature of the machine.
Abstract Recent research shows that anthropomorphizing of artificial intelligence (AI) technologies, such as virtual assistants or chatbots, affects users’ trust. This paper explores what an AI technology should look like at the interface to facilitate appropriate trust. It proposes an integrated framework based on related work, including Lee and See’s (2004) trust in automation model, Mollerup’s (2019) pretense design, as well as Clark and Fischer’s (2023) notion of depictions, for analysis and design of appearances of AI applications that pretend to be something different. Following the proposed framework, cases of AI applications are analyzed. They span a spectrum of appearances, from humans with open or specifc identities, through talking animals, to plants. Analyses suggest that appearances of AI applications should faithfully reflect the actual ability, characteristics, and design intent, in order to facilitate users’ trustworthiness assessments of the applications. Lastly, design implications and future research are discussed.
