What research says about AI interviews in qualitative data collection

The difference is clearest when AI interviews are compared with online surveys that use open-ended questions. Geiecke and Jaravel (2026) found that respondents in an AI interview provide about two and a half times as much information as people who answer the same question in an open text field. When independent reviewers compared the two directly, they rated the AI interviews as more informative in three out of four cases. Chopra et al. (2026) report a similar result. Per topic, AI interviews brought up about five times as many distinct reasons for respondents’ answers as a single open question, and the authors attribute this mainly to the follow-up questions the AI interviewer asks. The effect is older than the current generation of LLMs. In an earlier study by Xiao et al. (2020), a simple survey chatbot collected more than 5,200 responses that were more informative, more relevant and more specific than those from a standard web form.

This extra information is most useful where fixed answer categories capture only part of the reasoning behind an answer. Two people can give almost the same answer on a scale for very different reasons. Wuttke et al. (2026) found that supporters of different parties used different arguments even when their ratings on standardised scales were partly similar. Austin et al. (2026) confirmed this in an experiment with more than 2,200 participants. Adding AI follow-up questions to a standardised survey made the answers much longer and more specific.

AI interviews also hold up well against interviews led by humans. Geiecke and Jaravel (2026) asked researchers trained in qualitative methods to rate interview transcripts without knowing who had conducted them. Voice-based AI interviews received almost the same ratings as in-person interviews with human interviewers. Other studies likewise find no fundamental difference in quality between AI-led and human-led interviews (Wuttke et al., 2025; Guven et al., 2025). One reason may be that current research on AI interviewers designs its prompts around established quality criteria for qualitative interviewing (Helfferich, 2011).

The two kinds of interviewers do differ in how they run the conversation. Wuttke et al. (2025) observed that human interviewers sometimes steer respondents in a certain direction, while AI systems respond more neutrally to what people say. Respondents notice this too. In Wuttke et al. (2026), participants found AI interviews more motivating and engaging and felt more strongly that their own point of view had been captured. A conversation that feels more human does not automatically make a better interview, though. Cuevas et al. (2025) showed that an LLM-based interviewer made respondents more satisfied, yet their answers were no better than with a simple bot that only asked predefined questions. Overall, respondents see AI interviews positively. Across several studies, between 42 and 71 percent said they would rather talk to an AI again in a future interview. Only about 18 to 22 percent preferred a human interviewer, and the rest had no preference (Chopra et al., 2026; Geiecke & Jaravel, 2026).

The biggest advantage of AI interviews may be that they scale. Interviews led by humans take a lot of time and staff, while AI systems can run many interviews in parallel. Geiecke and Jaravel (2026) show that thousands of interviews can be completed within a few hours this way.

This matters when the goal is to measure how often themes and arguments appear and how they differ between population groups, in addition to identifying them. Chopra et al. (2026) found that with the same questionnaire, few new themes emerge after about 25 interviews. Reliable statements about group differences or about how common an argument is still require much larger samples, and AI interviews make such samples practical. In exploratory interviews, where each conversation can follow different questions, every additional round adds context and depth even after the themes are saturated.

Scale helps with the analysis as well. Once researchers have identified and defined the relevant themes in a sample of interviews, an LLM can code large amounts of interview material against these categories. Geiecke and Jaravel (2026) compared this kind of LLM-based coding with human coding. Agreement between the AI and human coders reached about 86 percent of the agreement between two human coders, and two different LLMs agreed with each other more closely than two human coders did. Wuttke et al. (2026) came to a similar result when they used the same approach to code more than 16,000 statement units from 571 interviews.

The results so far are promising. However, they do not mean that AI interviews work reliably regardless of how they are implemented.

The first issue is the dependence on model and prompt. Wuttke et al. (2025; 2026) show that switching from GPT-4o to GPT-5, as well as fairly small changes to the prompt, can change interviewer behaviour and sometimes make it worse. Different LLMs also behave quite differently. Panfilova et al. (2026) found differences in empathy, speed and in deciding when a follow-up question makes sense at all. This affects the quality of individual interviews. It also touches a central quality criterion of qualitative social research, which requires the research process to be transparent and verifiable by others (Bohnsack, 2005). If a model update or a small prompt change alters how the model behaves, the data collection becomes harder to reproduce.

A second limitation concerns the quality of the follow-up questions. Several studies show that AI systems sometimes move on to the next topic too early when an answer is interesting, surprising or unclear, instead of probing further (Wuttke et al., 2025; Lang & Eskenazi, 2025; Tirumala et al., 2025). Voice-based interviews add technical challenges. Transcription errors, emotion recognition and real-time conversation management all make them harder to use (Tirumala et al., 2025).

AI interviews therefore do not deliver reliable results on their own. Their quality depends on how carefully they are set up, including which model, prompt, transcription and analysis pipeline best fit the research question. In practice, AI interviews cannot simply replace human interviewers. They are a method in their own right and need to be adapted to the goals, sample and information needs of each study.