Imagine having a casual chat with an AI chatbot and leaving with a totally different opinion on a political issue that you felt strongly just 10 minutes ago. It sounds like a sci-fi movie, but it's already happening. New research shows that key AI models are highly effective at persuasiveness, and in some cases even more persuasive than humans. They use tone, evidence and personalization in ways that not only share facts, but also that allow them to coordinate their responses to individuals and subtly shake up opinions.
According to the report of Financial ExpressResearch conducted by the UK Institute of AI Security has found that AI models such as Openai's GPT-4, GPT-4.5, GPT-4O, Meta's Llama 3, Xai's Grok 3, and Alibaba's Qwen, in collaboration with universities, including Oxford and MIT, could influence political views in conversations of less than 10 minutes. Furthermore, the change in opinion was not fleeting. Most of the participants retained new views even after a month.
Researchers did not rely solely on the default behavior of AI. They tweaked these models using thousands of conversations on divisive topics such as health care funding and asylum policy. By rewarding outputs that match the desired persuasive style, AI was even more persuasive by adding personalized touches such as user age, political tendencies and referencing previous opinions. In fact, personalization increased persuasiveness by about 5% compared to general responses.
It may not be that big, but in the context of affecting public opinion, it is quite a bit. Political campaigns spend millions chasing a 1% swing on voter sentiment. The ability to get that shift in a few minutes on a massive scale is impressive and surprising. I think this is where the real debate begins. It's one thing for AI to sell new smartphones, but it's a completely different thing to tweak your stance on government policy.
The study also emphasized that AI persuasion is not limited to politics. Previous work by MIT and Cornell showed that these models could reduce beliefs about vaccine skepticism by engaging in conversations based on conspiracy theories, climate change denials, and personalized evidence. It sounds like a positive use case, but reinforces the fact that the same skill set can be applied in a less ethical way, such as spreading misinformation and promoting harmful ideology.
Interestingly, the persuasive power of AI also extends to commercial spaces. As Cornell's David Rand pointed out, chatbots could have a major impact on brand recognition and purchasing decisions. This feature can be an advantageous yet ethically grey revenue stream as tech companies like Openai and Google are trying to integrate advertising and shopping capabilities into AI assistants.
In my view, the real challenge is not only how compelling AI exists now, but how convincing it is in the next generation of models. Regulation and safeguards are extremely important, but so is the public's perception. If you know that the friendly chatbot you are talking about may be subtly manipulating your opinion, you may think twice before taking the words at face value. This technology is powerful and perhaps too powerful – and the world needs to step on carefully.
– end
