Algorithmic Reflectivity in Large Language Models: Methodological Study of Linguistic Structure in Mentalization

Mentalization is a core human capacity involving interpretation of one's own and others' behavior in terms of underlying mental states. Within the mentalization-based treatment (MBT) framework, it is described along multiple dimensions, integrating cognitive, affective, relational, and regulatory processes. Large language models (LLMs) can generate linguistically reflective discourse, raising the question of whether the formal linguistic structure of mentalization can bbe reproduced without experiential and affective processes.

This study aimed to assess whether an LLM can generate mentalization profiles structurally coherent with MBT dimensions and recognized as formally mentalizing by expert clinicians. Specifically, it evaluated the structural and argumentative coherence of model-generated profiles and the level of agreement among MBT-trained psychiatrists. It also examined the distinction between psychological mentalization and algorithmic reflectivity, conceptualized as a formally coherent but nonexperiential linguistic phenomenon.

A comparative, descriptive methodological design was adopted. A total of 50 human-model dialogues were generated under standardized conditions. At the end of each interaction, the model produced a narrative mentalization profile structured along MBT dimensions. Five MBT-trained psychiatrists independently evaluated each profile using anonymized materials. Ratings covered 4 canonical MBT axes and 2 additional higher-order dimensions-regulatory functions and synthetic profile. For each dimension, they assigned separate Likert-scale ratings for evaluative and argumentative coherence. Descriptive statistics were computed at dialogue level. Interrater reliability was assessed using intraclass correlation coefficients (ICCs), primarily ICC(3,k), with CIs. The association between evaluative and argumentative coherence was examined using Pearson correlation coefficients with 95% CIs, with Spearman rank correlation as a robustness check.

Across dimensions, mean scores ranged from 3.63 (SD 0.63) to 3.98 (SD 0.31), indicating moderate-to-high structural coherence. Interrater reliability was moderate to good, with ICC(3,k) values ranging from 0.60 to 0.84 in the evaluative domain. The highest scores were observed for implicit-explicit and synthetic profile, whereas internal-external showed the lowest mean. Score distributions showed limited variability and concentration in the upper range of the scale. Evaluative and argumentative coherence were strongly associated across all MBT dimensions, with Pearson r ranging from 0.826 to 0.967 (all P<.001) and Spearman ρ ranging from 0.866 to 0.991. Informal qualitative comments frequently described the outputs as linguistically organized and clinically interpretable, while also noting limited affective variability and contextual integration.

The findings suggest that LLMs can reproduce the formal linguistic structure of reflective discourse, generating outputs perceived by clinicians as coherent and clinically interpretable across multiple dimensions of mentalization. However, within the present evaluative framework, this performance appears more consistent with structural language regularities than with evidence of experiential, affective, or intentional processes. These results support using LLMs as research and training tools for studying the linguistic aspects of mentalization, while highlighting the continuing importance of human clinical judgment in contexts requiring emotional, relational, and contextual integration.