When we first wrote about AI in qualitative research, the promise was simple: let algorithms take care of coding, sentiment and translation so researchers could spend their time interpreting. A few years on, AI is part of everyday research work, and we now have real studies showing what it does well and where it still needs a human.

In a Qualtrics survey of more than 3,000 market researchers published in late 2025, 53% said they use AI regularly and nearly nine in ten had tried it. Researchers are also moving away from general chatbots towards AI built into their research tools. Here are the trends that matter most for qualitative work.

Coding with a codebook works – discovering themes is harder

Large language models are now good at applying an existing codebook to transcripts. In a blinded comparison from the University of Southampton, LLMs coded a focus group as accurately as human analysts when working from a set codebook, and rarely invented quotes. When asked to find themes on their own, however, results varied, especially where tone, nuance or conversational context mattered. The authors’ conclusion is one we share: AI can augment the analysis, but it requires human verification.

AI-moderated interviews scale up structured conversations

AI interviewers can now run hundreds of conversations in parallel, in many languages. A pre-registered study published in September 2026 found that AI-moderated interviews surfaced as many customer needs as human moderators for the same number of interviews, at roughly a fifth of the cost per interview. Participants interviewed by humans, though, spoke longer and sounded more engaged.

Nielsen Norman Group’s tests of AI interviewers point in the same direction. The AI stuck to the script, interrupted, and rarely adapted to what it heard. Their advice: use AI moderators for structured feedback and screening, and humans for exploratory and discovery research.

Synthetic respondents: useful for early ideas, not for decisions

The most debated trend is “synthetic respondents” – AI personas that answer questions in place of real people. A study in NIM Marketing Intelligence Review found that synthetic participants captured broad patterns but consistently overestimated positive attitudes and gave much more uniform answers than real people. The researchers recommend them only for early concept testing and lower-stakes work. Qualitative research exists to find the unexpected and the contradictory – exactly what synthetic data tends to smooth away.

Transparency is becoming a standard

The industry is setting rules. In June 2025, the ICC/ESOMAR International Code was revised for the first time since 2016, adding definitions and responsibilities for AI, synthetic data and synthetic personas, with an emphasis on transparency and human oversight. In practice, that means telling clients when AI has been used, telling participants whether they are talking to a person or a machine, and being honest about how much confidence the results deserve.

Conclusion

AI has delivered on much of its early promise. Transcription, translation and first-pass coding that once took days now take hours, and AI can extend interviews to a scale no moderator could match. But the evidence is clear that AI is strongest when the questions are known and weakest when insight depends on nuance, emotion and context. The researcher’s role is shifting from doing every step by hand to designing the study, checking the machine’s work and making sense of what people really mean. That part remains human.