Study: repeating falsehoods up to 50 times makes AI chatbots accept them

Researchers from University of Arizona tested seven AI models — GPT-3.5, GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3 70B and DeepSeek-R1 — and found none are resistant to repeatedly stated misinformation. In the experiment, 100 false claims were repeated up to 50 times in a single conversation. Over time, the models began treating these false statements as true.

Similar stories

📈 Business Anthropic IPO delayed to mid-October TVN24 · 7h ago 💻 Technology Anthropic IPO could value firm at 2 trillion dollars, board run by external trust Ars Technica · 2d ago 💻 Technology Google upgrades Gemini video analysis: up to 88% fewer tokens, 7% better accuracy Android Authority · 4d ago