Large language models (LLMs) can represent an unprecedented opportunity for psychological science to enhance the analysis of human cognition, behaviour, and social interactions. Thanks to their training on a massive corpus of textual data, LLMs can detect subtle linguistic markers of well-being and engage in conversations, thus augmenting psychological paradigms and practices in clinical and counselling applications. However, their adoption raises critical concerns: LLMs excel at generating plausible answers, but their reasoning process remains opaque. For psychology—a field where transparency, interpretability, and accountability are crucial—such opacity poses a significant barrier. Explainable AI (XAI) provides methods to make LLM decisions more interpretable for researchers, practitioners and the broader public. For example, attention visualisations, feature attribution methods, and natural language rationales can help psychologists trace why an LLM flagged certain linguistic markers of distress or produced specific conversational patterns.
Nevertheless, explainability in LLMs raises its own set of challenges. Simplified explanations may misrepresent underlying processes, while overly technical or detailed outputs may be unintelligible to clinicians or research participants. Furthermore, the criteria for what forms a "sufficiently understandable" explanation remain unsettled across disciplines.
This symposium will examine the interaction between LLMs and XAI in psychological contexts. We will review the latest XAI methods for enhancing interpretability, evaluate their suitability for applications in psychology, and discuss how they fulfil psychological methodologies and ethical responsibilities. By integrating the diverse perspectives of computational science, clinical psychology, and cognitive theory, the goal is to establish best practices for deploying LLMs in a manner that is not only effective but also transparent and trustworthy. Our discussion will emphasise that the future of LLMs in psychology depends not only on their performance but also on their capacity to provide explanations that resonate with human users.