The Use of Artificial Intelligence in Qualitative Research with Large Language Models: Potentials of Parameter Adjustment

Introduction

The integration of artificial intelligence (AI), particularly large language models (LLM) such as ChatGPT, LLaMA, or Mistral, into qualitative content analysis is a topic that is debated in the literature. While some authors highlight the potential of these tools, others raise concerns regarding their suitability. A central point of discussion is the response behavior of such models, which may vary even when identical prompts are used. Non-reproducibility and factual inaccuracies are cited as limitations. This study examines how targeted adjustments to model parameters can impact output in ways that enhance both reproducibility and analytical value.

Goals and Methods

The study examined three typical steps of qualitative content analysis, namely inductive category development, coding with a predefined codebook, and summarization of text segments. Two datasets of interviews were used to test a range of prompting strategies. In addition, systematic parameter adjustments were made, focusing on temperature, output diversity, and novelty preference, to assess their impact on the analytical process.

Results

Findings indicate that parameter settings substantially influence the quality and nature of LLM outputs. Lower temperature values and restricted output diversity produced more consistent and rule-based coding outcomes, which are favorable for reproducibility. Higher values fostered more associative and creative category development, better supporting exploratory approaches. Summarization tasks benefited from balanced settings that combined precision with interpretative openness. Overall, adaptive parametrization allowed for more predictable outputs and reduced methodological challenges such as inconsistent coding or analytically imprecise text interpretations.

Conclusions

The study suggests that qualitative research using LLM can benefit from systematic parameter adjustment. By calibrating parameters, it is possible to achieve a balance between methodological rigor and interpretative creativity. These findings provide an empirical foundation for more targeted and transparent use of AI in qualitative content analysis, offering new perspectives for enhancing both reliability and flexibility in AI-assisted research.