Show an AI chatbot the same false claim enough times, and it might start rating it as truer - but only if it's the right chatbot.
Researchers tested four language models - Gemma-3-4b-it, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, and GPT-5-nano - inside simulated social media feeds, repeating some statements while leaving others unseen within the same session. They gathered 336,000 truth, importance, sentiment, and interest ratings across 100 statements, 10 feed variants, and three replications, then ran regression and mixed-effects models to separate genuine belief shifts from simple familiarity effects. The four models split into distinct patterns: Gemma-3 showed a real illusory-truth effect, rating repeats as truer; Qwen2.5 showed only mere exposure, liking repeats more without judging them truer; GPT-5-nano barely moved and leaned mildly skeptical of repeated claims; and Llama-3.1 nudged up on truth while souring on other measures. The findings appear in an arXiv preprint titled "Illusory Truth or Mere Exposure? Model-Dependent Repetition Effects in LLM-Based Social Media Simulations" (arXiv:2609.36278), posted September 30, 2026.
This matters because researchers increasingly use these generative agent models to simulate how misinformation spreads across social networks, standing in for thousands of human users at a scale no lab study could match. If the models don't consistently replicate a well-documented human bias like the illusory-truth effect, conclusions drawn from those simulations - about which posts go viral or which interventions curb false beliefs - are only as sound as whichever model happens to be running them.
The paper's suggestion that Gemma-3-4b-it, the smallest model in the test, behaved most like a real person is either a nice underdog story or a reminder that bigger and newer doesn't automatically mean more human.