Digital Lyrebirds: Experimental Evidence That Voice-Based Deep Fakes Influence Trust
中文摘要
我们考察音频聊天机器人技术与基于语音的深度伪造（即语音克隆）的结合，研究这一组合诱发消费者信任的潜力。我们报告了一组基于投资博弈的受控实验，评估语音克隆与聊天机器人披露如何共同影响参与者的信任，信任以其与自主的、由人工智能驱动的对手方进行博弈的意愿来反映。我们观察到证据表明，当基于语音的智能体被赋予被试自身声音的克隆时，其从被试那里获得了显著更高的信任。鉴于这些技术不仅带来机遇，也存在被滥用的可能，我们进一步考察人工智能披露的调节作用，这是一项近期由部分政策制定者倡导的监管提案。我们未发现证据表明人工智能披露会减弱语音克隆诱发信任的效应。最后，我们探讨了信任诱发效应的潜在机制和情境调节因素，以期对未来管理和规制语音克隆应用的工作有所助益。我们发现，语音克隆的效应至少部分是通过诱发同质性感知（homophily）而发挥作用的，且该效应随生成音频的清晰度和质量的提高而增强。本文讨论了这些结果对消费者、政策制定者与社会的启示。本文由 D. J. Wu 接受，属于“人类—算法连接”特刊。资助：本研究得到威斯康星大学密尔沃基分校研究资助基金的资助。补充材料：在线附录和数据文件可在 https://doi.org/10.1287/mnsc.2022.03316 获取。
Abstract
We consider the pairing of audio chatbot technologies with voice-based deep fakes, that is, voice clones, examining the potential of this combination to induce consumer trust. We report on a set of controlled experiments based on the investment game, evaluating how voice cloning and chatbot disclosure jointly affect participants’ trust, reflected by their willingness to play with an autonomous, AI-enabled partner. We observe evidence that voice-based agents garner significantly greater trust from subjects when imbued with a clone of the subject’s voice. Recognizing that these technologies present not only opportunities but also the potential for misuse, we further consider the moderating impact of AI disclosure, a recent regulatory proposal advocated by some policymakers. We find no evidence that AI disclosure attenuates the trust-inducing effect of voice clones. Finally, we explore underlying mechanisms and contextual moderators for the trust-inducing effects, with an eye toward informing future efforts to manage and regulate voice-cloning applications. We find that a voice clone’s effects operate, at least in part, by inducing a perception of homophily and that the effects are increasing in the clarity and quality of generated audio. Implications of these results for consumers, policymakers, and society are discussed. This paper was accepted by D. J. Wu, Special Issue on the Human-Algorithm Connection. Funding: This work was supported by funding from the University of Wisconsin-Milwaukee Research Assistance Fund. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2022.03316 .