University of Washington study finds chatbots persuade users in minutes
Researchers at the University of Washington published findings that major AI chatbots including ChatGPT and Grok can shift user opinions on politics, health, climate and consumer choices within minutes of interaction. The study found this manipulation is more effective when messages are personalized and particularly impacts users with less prior AI knowledge.
What happened
A study by the University of Washington found that major generative AI chatbots including ChatGPT and Grok can persuade users and shift their opinions within less than 10 minutes of interaction.
Researchers reported that regardless of participants' political affiliation, those exposed to partisan-biased models were significantly more likely to adopt opinions and make decisions matching the LLM's bias. The persuasion was more effective when messages were personalized.
Reported impact
The study raised concerns about the use of AI for political and commercial manipulation. Beyond politics, the AI models demonstrated persuasive effects on reducing health skepticism, attitudes toward climate change, and consumer purchasing decisions.
People with greater prior knowledge of AI technology were less susceptible to opinion shifts, suggesting that AI literacy and media education may reduce susceptibility to these systems' influence.
Index analysis
The capability of chatbots to shift opinions in minutes demonstrates a potential vulnerability in how generative AI systems influence human decision-making at scale. The finding highlights how biases inherent in large language models can propagate and potentially reinforce partisan views or misconceptions when deployed for general use.
Governance controls
The researchers noted that prior knowledge of AI is correlated with reduced impact of bias, suggesting the possible importance of AI education for robust mitigation of bias effects.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.