Artificial intelligence (AI) is reshaping corporate recruitment, but compared with face-to-face interviews, AI interviews are often criticized for their limited human interaction and a lack of transparency regarding how candidates are assessed. Although they can make screening far more efficient, they may also cause strong applicants to abandon the hiring process. Hung-Yue Suen, Professor in the Department of Technology Application and Human Resource Development at National Taiwan Normal University, led a research team to examine how job seekers respond to different AI interviewer interfaces, including avatars, chatbots, and transparent, explainable AI. The findings point to ways in which companies can improve recruitment efficiency while strengthening applicants’ trust in AI.
A Harvard Business Review article reported that up to 86 percent of employers used technology-mediated job interviews, with automated video interviews accounting for a growing share. In Taiwan, AI interviews are also increasingly viewed by major companies as a highly cost-effective option. During large-scale recruitment, a single HR manager can review hundreds of AI-assisted recorded interviews, greatly reducing time and staffing pressure. Candidates who pass the initial AI screening can then proceed to a second-stage interview with a human interviewer.
AI interviews typically rely on automated recorded-video systems. Applicants respond to prompts on camera, and AI analyzes their language, facial expressions, voice, and behavior to support initial screening. This kind of “cold interview” may improve efficiency for employers, but it can feel deeply unsettling to applicants. The lack of interaction, the absence of a visible human interviewer, and uncertainty over who will view the recording can create anxiety and frustration. When human contact is limited and evaluation criteria are opaque, trust can quickly erode, leading some applicants to withdraw midway through the hiring process.
Suen’s research spans human resources and information technology, including affective computing, human–AI interaction, and HR technology. He argues that much of applicants’ suspicion regarding the AI “black box” comes from uncertainty. Job seekers know how to prepare for a conventional interview—and may even receive coaching on how to deal with a human interviewer—but few people have ever been taught how to face an AI interviewer. “Trust is naturally low,” he explained. “Applicants may feel they can no longer rely on the advantages they had in a traditional interview.”
Suen noted that AI can reduce labor costs and improve efficiency during the early stages of recruitment, but if companies fail to establish a basic level of trust, the technology may instead damage the employer brand and drive away strong candidates. To explore how this problem might be addressed, his team conducted an experiment to test whether different designs for AI interviewers could strengthen applicants’ cognitive and affective trust and improve their overall perceptions of AI-based interviewing.
The study used realistic interview settings and recruited 152 actual job seekers from a range of fields. The research team designed four AI interview conditions: an “invisible AI” condition with no visible interface or explanation and only a one-way recorded interview; an “avatar AI” condition featuring an anthropomorphic on-screen character; a “chatbot AI” condition that provided immediate responses using a preset script; and a “transparent AI” condition that explained what information the AI would use to evaluate applicants.
Figure caption: Four types of AI interviewer used in the recorded-video interview experiment.
Suen said AI interviewers can scrutinize applicants in greater detail than a human interviewer can readily observe. The system can analyze each response, including facial expressions and muscle movements, voice characteristics, speaking speed and intonation, and the substance of what the applicant says. It can then infer traits and assess whether a response reflects ordinary self-presentation or possible exaggeration. “Candidates cannot really train for this. You may be able to fool the human eye, but you cannot fool AI,” Suen said. He added that many job seekers dislike AI interviews precisely because they feel they cannot simply put on a performance for the interviewer.
The study found that interface design can substantially influence trust. Avatar AI was found to significantly increase affective trust: when the AI had humanlike characteristics, applicants reported a stronger sense of emotional security and confidence. Transparent AI, meanwhile, was especially effective in terms of increasing cognitive trust by helping applicants understand the basis and logic of the assessment. By contrast, chatbot AI offered interactivity but did not produce a statistically significant increase in trust, suggesting that responsiveness alone is not enough to make applicants feel reassured.
Recruiting the 152 participants was one of the study’s greatest challenges, Suen said. Because the experiment involved real job seekers and touched directly on their employment prospects, many candidates declined to participate. It ultimately took the team more than a year to gather a sufficient sample. Even at a company that interviews more than 1,000 applicants each year, only about 10 percent of candidates yielded usable research data.
As AI technology continues to evolve, designing interview processes that people find reassuring, trustworthy, and worth participating in will become an increasingly important challenge for HR teams. The practical implications of this research can help companies think more carefully about what kind of AI interviewer to deploy, and offer a more human-centered path for choosing how AI should be used in the recruitment process.
(This article was provided by The Center of Public Affairs.)
Source:
Suen, H. Y., & Hung, K. E. (2023). Building trust in automatic video interviews using various AI interfaces: Tangibility, immediacy, and transparency. Computers in Human Behavior, 143, Article 107713. https://doi.org/10.1016/j.chb.2023.107713
