Automating Adult Attachment Interview Classification Using Large Language Models

This project investigates whether large language models (LLMs) can be used to support the automated classification of Adult Attachment Interview (AAI) transcripts in a way that is theoretically aligned with established AAI coding principles. The Adult Attachment Interview is widely regarded as a gold-standard measure of adult attachment, but its reliance on expert judgment makes it time-intensive and difficult to scale. Recent advances in large language models provide new opportunities to analyse complex psychological texts at the level of discourse and reasoning, rather than relying solely on surface linguistic features. However, their application to AAI transcripts remains largely unexplored. The study consists of two stages. In Study 1, an instruction-following large language model deployed within a secure Azure environment will be used to perform zero-shot classification of English AAI transcripts into four attachment categories: Secure, Dismissing, Preoccupied, and Unresolved. The model will be prompted to base its judgments on core AAI coding principles, such as coherence of discourse, balance of evaluations of attachment figures, and lapses in reasoning around loss or trauma. Model outputs will be compared against expert-coded attachment classifications. In Study 2, the same model will be fine-tuned using a subset of labelled transcripts to examine whether task-specific alignment improves classification performance and the quality of model-generated explanations. Performance will be evaluated using standard classification metrics, alongside qualitative assessment of the model’s reasoning. All data processing and analysis will take place within a secure institutional computing environment, and no data will be stored or processed locally. This project aims to contribute to methodological discussions around the use of large language models in clinical and developmental psychology, and to explore their potential role in supporting, rather than replacing, expert-based attachment assessment.