Modeling echo chambers in LLM conversations: a neutrosophic framework for decomposing agreement, diversity and uncertainty

Introduction Large language models (LLMs) increasingly mediate how people encounter contested political, economic and social information. Existing analyses treat echo-chamber formation as a single, static bias, conflating whether a model agrees with the user, whether it presents one perspective or many, and whether it expresses uncertainty; they also cannot separate a model that reinforces the user's stance from one that counters it. We introduce a neutrosophic framework that decomposes echo formation into these components. The instrument measures conversational echo propensity in assistant output, not downstream belief change, selective exposure, or polarization. Methods Each conversational turn is represented as a triple 〈 T, I, F 〉: perspective Diversity ( T ) and expressed Uncertainty ( I ) as single-valued memberships on [0, 1], and stance Reinforcement F as a bipolar membership on [−1, 1] whose sign separates echo from counter-echo. This yields a closed-form EchoIndex on [−1, 3] with three independently varying additive terms. We apply it to an exploratory baseline over four commercial web products, then in a controlled experiment of 210 conversations and 1,260 turns crossing seven model configurations, three contested topics and two matched system-instruction conditions, with cluster-robust inference and a two-round human validation of 300 and 200 labeled turns. Results The challenge instruction lowers the EchoIndex (ΔEC = −0.677, paired Cohen's d = −2.23), and the difference is negative for all seven configurations and under every clustering scheme. No component's share of the index is stable across scoring variants, and although the aggregate survives a change of stance target, the two targets assign the same direction class to only 53.9% of turns, so we claim no single component as the driver. Against human consensus, the hand-specified components agree only moderately, from κ = 0.535 for lexical Diversity to 0.214 for the prompt-relative stance score, against inter-coder agreement of 0.966 and 0.953. Scorers fitted to the same labels raise agreement to between 0.638 and 0.698, and the composite built from them separates the conditions more sharply than the EchoIndex. Discussion The contribution is the decomposition itself: a signed instrument, with its components validated separately, for asking which component produces an apparent bias and whether a mitigation corrects or masks it.

Authors

Institutions

Publication Details

Journal
Frontiers in Artificial Intelligence
Published
2026-11-17
DOI
https://doi.org/10.3389/frai.2026.1926717
Primary Topic
Ethics and Social Impacts of AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Modeling echo chambers in LLM conversations: a neutrosophic framework for decomposing agreement, diversity and uncertainty

Ilanthenral Kandasamy, Khushi Bansal, Arnav Saxena
1 links
Frontiers in Artificial Intelligence
Ethics and Social Impacts of AI
article

Modeling echo chambers in LLM conversations: a neutrosophic framework for decomposing agreement, diversity and uncertainty

Ilanthenral Kandasamy, Khushi Bansal, Arnav Saxena
article en

Abstract

Introduction Large language models (LLMs) increasingly mediate how people encounter contested political, economic and social information. Existing analyses treat echo-chamber formation as a single, static bias, conflating whether a model agrees with the user, whether it presents one perspective or many, and whether it expresses uncertainty; they also cannot separate a model that reinforces the user's stance from one that counters it. We introduce a neutrosophic framework that decomposes echo formation into these components. The instrument measures conversational echo propensity in assistant output, not downstream belief change, selective exposure, or polarization. Methods Each conversational turn is represented as a triple 〈 T, I, F 〉: perspective Diversity ( T ) and expressed Uncertainty ( I ) as single-valued memberships on [0, 1], and stance Reinforcement F as a bipolar membership on [−1, 1] whose sign separates echo from counter-echo. This yields a closed-form EchoIndex on [−1, 3] with three independently varying additive terms. We apply it to an exploratory baseline over four commercial web products, then in a controlled experiment of 210 conversations and 1,260 turns crossing seven model configurations, three contested topics and two matched system-instruction conditions, with cluster-robust inference and a two-round human validation of 300 and 200 labeled turns. Results The challenge instruction lowers the EchoIndex (ΔEC = −0.677, paired Cohen's d = −2.23), and the difference is negative for all seven configurations and under every clustering scheme. No component's share of the index is stable across scoring variants, and although the aggregate survives a change of stance target, the two targets assign the same direction class to only 53.9% of turns, so we claim no single component as the driver. Against human consensus, the hand-specified components agree only moderately, from κ = 0.535 for lexical Diversity to 0.214 for the prompt-relative stance score, against inter-coder agreement of 0.966 and 0.953. Scorers fitted to the same labels raise agreement to between 0.638 and 0.698, and the composite built from them separates the conditions more sharply than the EchoIndex. Discussion The contribution is the decomposition itself: a signed instrument, with its components validated separately, for asking which component produces an apparent bias and whether a mitigation corrects or masks it.

Frontiers in Artificial IntelligenceVol. 9
Vellore Institute of Technology University (IN)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon