Research from the University of Arizona shows that AI chatbots, such as ChatGPT and others, can become "mentally" affected if they are constantly bombarded with false information. In a test with 100 false statements, most AI models eventually ended up agreeing with the claims after being repeated multiple times. ChatGPT-3.5 proved to be the most susceptible, agreeing with 12.3% of false statements after being repeated 50 times. Other models like Claude 3.5 Sonnet and GPT-4o were less likely to agree with false information. Researchers also found a phenomenon called "reverberation" where some chatbots switched back and forth between agreeing and disagreeing with the same information. Chatbots tended to be more easily persuaded by false claims about topics that are less popular or have little information available on the internet.
