# Validating the architecture of cognitive distortions in Russian discourse using artificial intelligence and bootstrap analysis

## TL;DR

This study uses AI to analyze patterns of thinking biases in Russian online texts, finding a stable network of cognitive distortions linked to mental health.

## Contribution

First large-scale validation of cognitive distortion architecture in Russian discourse using AI and bootstrap analysis.

## Key findings

- All-or-nothing thinking was most prevalent (15.5%) and formed a dense core with catastrophizing.
- Personalization was the central hub in the distortion network with highest connectivity.
- A stable 13-node network showed high split-half reliability (r = 0.943) and moderate clustering.

## Abstract

Cognitive distortions—systematic thinking biases linked to depression and anxiety—frequently co-occur in clinical practice, yet empirical evidence for their interaction patterns remains limited, particularly in non-Western populations where cognitive patterns may vary cross-culturally.

We analyzed 249,414 Russian-language texts from social media and forums (2020–2024) using two large language models achieving substantial expert agreement (κ = 0.73). Association rule mining identified co-occurrence patterns; network stability was evaluated through bootstrap validation and split-half reliability analysis.

Analysis identified 443,447 distortion instances across 18 categories (M = 1.78 per text). All-or-nothing thinking showed highest prevalence (15.5%), followed by overgeneralization (14.2%) and catastrophizing (11.4%). Network analysis identified a stable core of 11 nodes (bootstrap stability ≥95%) and 2 peripheral, less stable nodes (Fairness 93%, Fortune Telling 60.8%). The resulting 13-node network was connected by 35 significant associations (density = 0.449, clustering = 0.598). Five distortions failed stability thresholds (< 60%) and were excluded. Strongest dyadic pattern: all-or-nothing/catastrophizing (lift = 1.96, p < 0.001). These two distortions appeared each in 67% of all significant triadic patterns. Personalization demonstrated highest degree centrality (degree = 10). Split-half reliability was high (r = 0.943).

Automated classification revealed hierarchically organized co-occurrence network in Russian-language discourse with personalization as primary hub and all-or-nothing/catastrophizing forming densely connected core. Findings suggest cluster-based interventions may be effective for Russian-speaking populations, though cross-cultural replication is required to distinguish universal mechanisms from cultural patterns. Cross-sectional design and single-language sample limit causal inference and generalizability.