How do large language models (LLMs) address the limitations of traditional supervised learning for qualitative analysis?

LLMs address the limitation of relying on labeled training data by acting as general-purpose classifiers that use their statistical representation of language and simple prompts instead of supervised training. They also offer flexibility through in-context learning, allowing researchers to classify qualitative data with zero, one, or a few examples and to adapt the classification logic to specific domains.

Traditional supervised learning for classification requires a model to be trained on manually labeled data, and the approach is rigid because it is designed for narrowly defined tasks. This reliance on labeling restricts the amount of training data and makes qualitative researchers skeptical of using such tools. LLMs overcome this by functioning as general-purpose classifiers: rather than being trained on labeled examples for each task, they leverage their pretrained statistical representation of language to classify text based on explicit instructions given in a prompt. This enables them to work with both structured text records and unstructured data such as interviews, archival documents, and observations, which are central to qualitative analysis. Their ability to generalize across domains and learn new information directly from the prompt, known as in-context learning, supports zero-shot, one-shot, and few-shot classification. This flexibility reduces the need for labeled data because only a few labeled examples are needed in the prompt, facilitates quick experimentation with different classification schemes, and enables domain adaptation by tuning classification logic to the nuances of a specific field. At the same time, LLM-based classification is sensitive to prompt wording and example order, and the number of examples is limited by context size, so researchers should validate outputs and treat LLMs as complementary analytical tools rather than infallible classifiers.