From Echo Chambers to Echo Machines? Endorsement, Evidence, and the Certainty of Belief in Conversations With AI
Abstract
Randomized trials show that conversations with large language models (LLMs) can reduce conspiracy belief and, more modestly, political polarization. Yet the same systems can increase conspiracy belief when instructed to, and clinical reports describe chatbot conversations accompanying the escalation of implausible convictions. Prevailing accounts attribute both outcomes to the evidence a model supplies. We note that most experiments confound evidence about the world with endorsement of the person, and that the clearest attempt to separate them favors evidence. Drawing on research on attitude certainty, self-validation, and responses to flattery, we distinguish the position of a belief from the certainty with which it is held and formalize γ, an effect of endorsement on certainty at fixed factual content, interpretable as a validation gain. If higher certainty lowers responsiveness to later evidence, the account permits a conditional outcome: beliefs that become less extreme yet more rigid, without any misinformation. We specify two critical experiments, a factorial test of facts and endorsement that includes a source known to approve everyone and an endorsement-first design with counterevidence held constant, plus an articulation control. Conversational AI becomes a test bench for an old question: can social approval make subsequent evidence matter less?
Related
- Understanding, Confidence and Usefulness of Effect Size and Certainty of Evidence: A Cross-Sectional Survey Among Australian Dietitians
- Evidence Mapping, Certainty Appraisal, and Acupoint Prescription Pattern Mining of Acupuncture for Poststroke Dysphagia: An Overview of Systematic Reviews.
- Autologous Nanofat Injection for Cutaneous Scars: A PRISMA 2020 Systematic Review of Clinical Outcomes and Certainty of Evidence
- Associations between pretreatment systemic immune-inflammation index and tumor-specific prognosis in breast and gynecological cancers: an updated systematic review, meta-analysis, and certainty-of-evidence assessment
- Evaluating Certainty of Evidence Assessment in Systematic Reviews of Yoga for Health: A Meta‐Research Study
BibTeX
@misc{le2026from,
title = {{From Echo Chambers to Echo Machines? Endorsement, Evidence, and the Certainty of Belief in Conversations With AI}},
author = {Olivier LE BON and Xavier Noël},
howpublished = {PsyArXiv},
year = {2026},
month = oct,
doi = {10.31234/osf.io/7et5g_v1},
url = {https://osf.io/preprints/psyarxiv/7et5g_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - From Echo Chambers to Echo Machines? Endorsement, Evidence, and the Certainty of Belief in Conversations With AI
AU - Olivier LE BON
AU - Xavier Noël
PY - 2026
DA - 2026/10/23
PB - PsyArXiv
DO - 10.31234/osf.io/7et5g_v1
UR - https://osf.io/preprints/psyarxiv/7et5g_v1/
AB - Randomized trials show that conversations with large language models (LLMs) can reduce conspiracy belief and, more modestly, political polarization. Yet the same systems can increase conspiracy belief when instructed to, and clinical reports describe chatbot conversations accompanying the escalation of implausible convictions. Prevailing accounts attribute both outcomes to the evidence a model supplies. We note that most experiments confound evidence about the world with endorsement of the person, and that the clearest attempt to separate them favors evidence. Drawing on research on attitude certainty, self-validation, and responses to flattery, we distinguish the position of a belief from the certainty with which it is held and formalize γ, an effect of endorsement on certainty at fixed factual content, interpretable as a validation gain. If higher certainty lowers responsiveness to later evidence, the account permits a conditional outcome: beliefs that become less extreme yet more rigid, without any misinformation. We specify two critical experiments, a factorial test of facts and endorsement that includes a source known to approve everyone and an endorsement-first design with counterevidence held constant, plus an articulation control. Conversational AI becomes a test bench for an old question: can social approval make subsequent evidence matter less?
ER -