Chatbots can be extremely effective political persuaders, especially when they produce a high volume of factually incorrect information.
Artificial intelligence (AI) chatbots like ChatGPT and Claude are becoming a major source of information, guiding users on everything from health to politics.
About one in eight (13%) of all eligible UK voters reported using AI chatbots to seek out information directly relevant to their electoral choice in the 2024 general election.
But should you allow a chatbot to influence your political opinion? A new study published in the journal Science found that the more persuasive chatbots are on political issues, the more likely they are to provide factually inaccurate information.
Researchers conducted three large-scale survey experiments involving nearly 77,000 participants from the UK who conversed with 19 different large language models (LLMs) on 707 political issues.
After each interaction, participants reported their level of agreement with statements expressing a particular political opinion relevant to the UK on a 0-100 scale. A total of more than 466,000 AI-generated fact-checkable claims were included in the analysis.
The research team also tested multiple prompting strategies and post-training methods, evaluating how each “lever” affected the persuasive impact and factual accuracy of LLMs.
The study found that model size and personalization – providing the LLM with information about the user – had only a minor impact on the persuasiveness of the models.
Post-training techniques and prompting strategies, on the other hand, resulted in dramatic increases in persuasiveness of 51% and 27%, respectively.
The models were found to be most persuasive when their arguments included a high volume of factual claims. However, this is where the researchers observed a concerning trade-off between persuasion and accuracy.
Post-training and information-focused prompting, the levers that made LLMs more persuasive, systematically caused them to produce less factually accurate information.
In other words, the models that produced the most factual claims also tended to be less accurate on average. For example, GPT-4o made accurate claims 62% of the time when prompted to use information versus a different prompt (78%).
LLMs are convincing, but are we convincible?
The findings come amid rising concerns that AI may disrupt the integrity of democratic processes by being deployed in misinformation campaigns and foreign interference.
A recent study suggests that AI can pass checks designed to detect automated responses in opinion polls, potentially swaying election results and allowing foreign adversaries to meddle in politics.
However, when it comes to LLMs’ power of persuasion in the real world, they face a critical bottleneck.
The new study says that despite lengthy, information-dense conversations with LLMs being highly persuasive, the extent to which people will voluntarily sustain cognitively demanding political discussions with AI systems outside of a survey context remains unclear.
Researchers suggest that many people lack awareness or interest in politics and are being bombarded with competing demands on their attention.
Unlock exclusive Cybernews content on YouTube
Your email address will not be published. Required fields are markedmarked