GPT-4o interview assistant raises 5 ethical red flags in study

A new study published at CHIWORK '26 examines the ethical pitfalls of using large language models like GPT-4o to generate real-time follow-up questions in semi-structured interviews. Researchers from Penn State and Syracuse University designed a Wizard-of-Oz setup where a human co-interviewer could selectively relay or edit AI-generated questions, simulating deployment without full automation. Across 17 interviewers with varied qualitative research experience, participants identified five critical concerns: harmful or discriminatory language in AI outputs, diminished respect for interviewees due to divided attention and missed nonverbal cues, technology-based participation inequality (e.g., data access barriers), unclear accountability when harms occur, and serious privacy risks when the AI listens, records, or transcribes sensitive content.

The study translates these concerns into concrete design and governance implications. For tech professionals building or deploying AI assistants in HR, journalism, or research contexts, the findings highlight that even human-in-the-loop systems introduce new failure modes – not just technical errors but social and ethical harms. The authors argue for explicit responsibility frameworks, real-time bias detection, consent-aware data handling, and interface designs that keep the interviewer fully engaged. As AI interview assistants move from labs to real-world use, this research provides a timely cautionary framework for responsible deployment.