JMIR Publications

ABSTRACT

Background:

Mentalization is a core human capacity involving the interpretation of one’s own and others’ behavior in terms of underlying mental states. Within the Mentalization-Based Treatment (MBT) framework, this capacity is described along multiple dimensions integrating cognitive, affective, relational, and regulatory processes. Large language models (LLMs) have recently shown an ability to generate linguistically reflective discourse, raising questions about whether the formal linguistic structure of mentalization can be reproduced independently of experiential and affective processes. This study investigates whether LLM outputs can be systematically evaluated as reflecting the linguistic structure of mentalization without implying psychological mentalization or theory of mind.

Objective:

The aim of this study is to assess whether a large language model can generate outputs that are structurally coherent with established MBT dimensions, and to determine the extent to which such outputs are recognizable as formally mentalizing by expert clinicians. The study introduces and operationalizes the concept of algorithmic reflectivity, defined as a formally coherent but non-experiential linguistic phenomenon.

Methods:

A comparative, descriptive methodological design was adopted. Fifty dialogic interactions between a large language model and human participants were generated under standardized conditions. At the end of each interaction, the model produced a narrative mentalization profile structured along MBT dimensions. Five psychiatrists with formal MBT training independently and blindly evaluated all profiles. Evaluations used 5-point Likert scales assessing (1) evaluative coherence, (2) argumentative coherence, and (3) global quality across the MBT dimensions. Interrater reliability was estimated using the intraclass correlation coefficient (ICC[3,1]). Descriptive statistics were used to summarize score distributions and variability.

Results:

Across all dimensions, mean scores ranged from 3.63 to 3.98, indicating moderate to high structural coherence. Interrater reliability was substantial to high, with ICC values ranging from 0.60 to 0.84. The highest scores were observed for dimensions related to explicitness, synthesis, and self–other differentiation, while lower scores were observed for integration between internal states and external context. Qualitative comments consistently described the outputs as linguistically organized and clinically interpretable, but affectively neutral and weakly contextualized. No evidence of experiential grounding, affective modulation, or intentional agency was observed.

Conclusions:

The findings indicate that LLMs can reliably reproduce the formal linguistic structure associated with mentalization as defined by MBT, generating outputs that expert clinicians recognize as structurally coherent. However, this capacity reflects algorithmic reflectivity rather than psychological mentalization: a form of linguistic coherence without experiential, affective, or relational grounding. The study supports a clear conceptual distinction between mentalization as a psychological function and its discursive structure as a linguistic phenomenon. These results suggest that LLMs may serve as methodological tools for research and training on reflective language, while remaining unsuitable for unsupervised clinical application.