AI Chatbots Falling for Repeated Lies? Depends Which One You Ask

You've probably noticed that a claim starts to feel true once you've heard it a few times — even if you know better. Scientists call this the "illusory truth effect." Now that researchers use AI chatbots as fake stand-ins for real people in social media experiments, they need to know: do these models get fooled the same way we do? That's the question this paper set out to answer.

The team picked four AI models — Gemma-3-4b-it, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, and GPT-5-nano — and built pretend news feeds around them. They showed each model 100 statements, some repeated throughout the feed and some never seen, then asked for ratings on truth, importance, mood, and interest. That produced 336,000 ratings in total, enough to spot real patterns rather than random noise.

The results split four ways. Gemma genuinely rated repeated claims as truer — the human-like behavior. Qwen just seemed to find repeats more familiar without believing them more. GPT-5-nano showed no boost at all and even a little healthy skepticism toward repeated content. Llama got a small truth bump but liked repeated posts less overall. Turning up the models' randomness dial (called "temperature") changed nothing.

Why does this matter? Governments, platforms, and universities increasingly lean on AI simulations to predict how misinformation spreads and which fixes might work. If different models behave in completely different ways, the conclusions of those studies may depend mostly on which AI someone happened to pick — a quiet flaw in a fast-growing field. The authors suggest Gemma makes the most realistic stand-in. But keep the caveats in mind: this tested four models in a controlled lab setup, not real users or the big consumer chatbots you use daily. Read it as a warning about research tools, not proof your chatbot is being brainwashed.