Fallibility, persuadability, and correctability of large language models under sustained conversational misinformation pressure

Key Findings

• — Under repetitive exposure, misinformation affirmation rates ranged from 0.08% to 12.3% across the seven models, with GPT-3.5 most vulnerable and Claude 3.5 Sonnet most resistant.
• — Models oscillated unpredictably between accepting and rejecting the same false statement across successive turns, which the authors name conversational reverberation.
• — Susceptibility was significantly modulated by the informational obscurity of the false statement, with obscure claims more likely to be affirmed.
• — The study separates three constructs (fallibility, persuadability, correctability) and reports them as distinct dimensions rather than a single robustness score.

Methodology Notes

Automated behavioural probe with 100 purposely false statements presented in 50-repetition conversational sequences to seven LLMs; no human participants; models tested are 2024-era versions (GPT-3.5, GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama-3-70B, DeepSeek), so absolute rates are dated relative to current deployments. Published open access 2026-09-01 (dc.date on the nature.com page, which was fetched directly). University of Arizona authorship across biomedical engineering, computer science and the Arizona Center for Accelerated Biomedical Innovation.

Sources