Human-in-the-Loop: Qualitative Research in the Age of AI
António Pedro Costa, University of Aveiro (Portugal)
Researcher at the Research Centre on Didactics and Technology in the Education of Trainers (CIDTFF), Department of Education and Psychology, University of Aveiro, and collaborator at the Laboratory of Artificial Intelligence and Computer Science (LIACC), Faculty of Engineering, University of Porto.
The integration of Artificial Intelligence (AI), essentially Generative Artificial Intelligence (GenAI), into qualitative research is redefining the way we produce knowledge. Generative tools, language models, and AI agents have transformed previously time-consuming tasks such as transcription, coding, or categorisation into automatic and scalable processes. However, this methodological acceleration raises a central challenge: how can we ensure that interpretation, reflexivity, and human meaning are not lost in automation?
As Costa et al. (2025a) argue in “AI as a Co-researcher in the Qualitative Research Workflow: Transforming Human-AI Collaboration”, AI can be understood not only as a tool but as a co-researcher. The AbductivAI model proposes precisely such a collaboration between humans and AI agents, integrating deductive, inductive, and abductive reasoning through Chain-of-Thought prompting. This approach does not eliminate the human role but rather reinforces it by requiring the researcher to act as a reflective agent, responsible for validating and interpreting the results produced by AI. The notion of human-in-the-loop takes on an epistemological meaning here. Humans are not only there to supervise the algorithm but also to co-produce knowledge, mediating tensions between automatic inference and contextual interpretation. The researcher becomes a curator of meaning, ensuring ethical and methodological rigour in a “hybrid intelligence” ecosystem.
This idea of collaboration extends to the proposal for Mixed-Methods & AI for Methodological Literature Reviews – MIXAI (Costa et al., 2025b), which demonstrates how combining AI with mixed methods enables richer and more systematic methodological reviews. In the MIXAI model, human researchers and AI work in triangulation: bibliometric (quantitative) analysis is complemented by framing (qualitative) analysis using Chain-of-Thought prompting in models such as ChatGPT and Gemini. The result is a quant-QUAL integration, in which AI expands analytical capacity, but the researcher retains interpretative and ethical decision-making. More than just automating, AI highlights new skills that researchers need: AI literacy, critical thinking, ethical awareness, and interdisciplinarity. These skills shape the new profile of the reflective-digital researcher, capable of understanding the potential and limitations of generative systems.
If AbductivAI emphasises collaboration and MIXAI emphasises integration, the RUF – Reflexive Uncertainty Framework (Costa & Bem-Haja, 2025c) adds a fundamental layer: uncertainty as methodological evidence. Instead of treating the ambiguities of probabilistic models as errors, RUF proposes viewing them as opportunities for reflection and dialogue. During data analysis, areas of ambiguity, where coding probabilities converge or model justifications oscillate, are interpreted as moments for reflection. This is where humans must intervene: questioning, justifying, and making explicit the limits of “automatic interpretation”, ensuring ethical boundaries through the responsible use of GenAI. RUF turns uncertainty into a productive space for co-analysis, strengthening epistemological transparency and human agency in AI-mediated research.
The three models, AbductivAI, MIXAI, and RUF, converge on the same premise. AI does not replace the researcher; it expands their ability to see, but only if the human remains in the loop. The real methodological innovation is not in automating analysis but in reimagining the relationship between humans and AI as a critical, ethical, and reflective partnership. Keeping humans in the loop is thus an epistemological requirement of qualitative research. It means recognising that interpretation is always situated, affective, and socially mediated, something that no probabilistic model can replicate. By accepting uncertainty, practising reflexivity and engaging in dialogue with generative systems, the researcher ensures that the ‘guardian’ of qualitative research remains human, even in the age of AI.
References
Costa, A. P., Bryda, G., Christou, P. A., & Kasperiuniene, J. (2025a). AI as a Co-researcher in the Qualitative Research Workflow: Transforming Human-AI Collaboration. International Journal of Qualitative Methods, 24. https://doi.org/10.1177/16094069251383739
Costa, A. P., Burneo, P., & Kasperiuniene, J. (2025b). Mixed-methods & AI for Methodological Literature Reviews. The Qualitative Report, 30(10), 4515–4540. https://doi.org/10.46743/2160-3715/2025.8434.
Costa, A. P., & Bem-Haja, P. (2025c). Reflexive Uncertainty AI for Qualitative Data Analysis. Proceedings of the 28th European Conference on Artificial Intelligence (ECAI 2025), in press.
