Answers are grounded in this paper and cite their sources.
This exploratory study examined the possibilities and limitations of using generative AI as a supplementary tool for qualitative coding of written reflections produced by students in Swedish special education programmes. The empirical material comprised 99 responses to a debate article on inclusion and special educational perspectives. A theory-informed coding framework distinguished compensatory, relational and dilemma perspectives as non-exhaustive, potentially overlapping ideal types in naturally occurring texts. Two human raters independently coded a stratified random subsample of 20 texts, with ChatGPT as an additional coder using a specified prompting procedure. Human–human agreement was 80% (Cohen’s kappa = 0.70); agreement between each human rater and ChatGPT was 65% (kappa = 0.46 and 0.42). Under the tested configuration, ChatGPT identified no compensatory cases and tended to interpret peripheral forms of inclusion as dilemma-oriented. A post hoc prompt refinement, informed by errors in the same subsample and applied only to six compensatory cases, changed several outputs but provided no independent test of improved performance. Findings therefore concern this model-prompt-data configuration rather than large language models’ inherent capabilities. Agreement, systematic disagreement and prompt sensitivity can inform methodological reflection, underscoring the need for transparent procedures, held-out validation and human interpretive judgement in normatively complex qualitative analysis.
No takes yet. Share an insight, caveat, or question.
Karlsudd et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: