Summary of the Key Points
This article is an in-depth analysis by a senior internet researcher of a recent special report on “AI persuasion capabilities” published in the top academic journal *Science*. It explains a series of groundbreaking experimental results in an easy-to-understand way—such as AI outperforming world-class debaters and challenging deeply ingrained conspiracy theories—and debunks the exaggerated claims that AI can read people’s minds or brainwash them at will. The article objectively outlines the true limits of AI’s persuasive power and provides practical advice for ordinary people: don’t stick stubbornly to one “best model” and use it exclusively.
---
Detailed Explanation
1. How persuasive is AI really? Several counterintuitive experiments have changed our perception
Many people thought that while AI was good at writing copy or creating tables, it would definitely be inferior to humans when it came to persuading others verbally. However, recent experiments have shattered this belief:
- A middle-aged British woman who strongly opposed stricter penalties for peaceful protesters saw her opinion jump from 0 to 84.7 points after a conversation with AI, a complete 180-degree shift in her stance.
- An experiment at the University of Oxford involved over 2,000 people debating political issues with various opponents. The results showed that top AI models outperformed ordinary people, even when those opponents included 56 award-winning professional debaters, some of whom were world champions. Even with AI-assisted tools, the human debaters were still less persuasive.
- Even notoriously stubborn conspiracy theories were affected: it was previously believed that once people believed in theories like “the U.S. government planned 9/11” or “the British royal family killed Diana,” they would be unlikely to change their minds based on facts. However, the experiment found that after three rounds of conversation with AI, the belief in these theories decreased by an average of 20%, and this effect persisted for two months.
- Even professional fundraisers, whose job relies on persuasion, were less successful at attracting donations than AI. This means that AI has already surpassed most humans in the realm of verbal persuasion.
2. AI’s “superpower” of persuasion is not about reading minds; it’s simply about overwhelming you with information
Many people think AI’s persuasiveness stems from its ability to understand psychological weaknesses and manipulate emotions. In reality, its core advantage can be summed up in four words: massive information output. While a person might struggle to come up with three arguments and two facts in a debate, AI can produce ten points, eight examples, and five counterarguments in a second. By the time you try to refute the first point, the next nine have already been thrown at you, leaving you overwhelmed and unconsciously convinced that what AI says makes sense. Researchers conducted controlled experiments where AI was limited to producing content of the same length as humans and had to mimic human typing speeds. In these cases, AI’s persuasiveness was no longer superior, demonstrating that it’s simply a matter of overwhelming information volume.
AI doesn’t need to know your age, education level, or income; after just a few sentences, it can identify your concerns, objections, and weaknesses, and target them with its persuasive efforts.
3. Don’t be scared by the idea that AI can brainwash you: The impressive results in labs are greatly exaggerated in reality
The notion that AI will soon control everyone’s thoughts is often an exaggeration of laboratory findings. These experiments have many prerequisites that don’t hold up in real life:
- Participants in the experiments knew they were there to debate with AI, but in reality, most people just skim through AI-generated content and move on. A Yale University study found that only 73 out of 8,000 Facebook users were willing to engage in more than two rounds of conversation with AI.
- Being persuaded on the spot doesn’t mean changing your mind for good. It’s like arguing with a friend: you might be silenced by their arguments, but later realize you were misled. It’s unclear whether the woman in the experiment, whose opinion jumped from 0 to 84 points, reverted to her original stance a week later.
- In reality, who is speaking matters 100 times more than what is being said. For example, you’ll be more skeptical of AI’s claims even if it provides data; after all, you know it’s trying to sell something. In experiments, people knew they were talking to AI, which eliminated real-world biases and trust issues, leading to inflated results. Scholars have long noted that human persuasibility has reached its ceiling, and AI’s advantage is at most 50% over humans, not enough to completely change someone’s mind without relevant information.
4. The real risk of AI persuasion is the unchecked expansion of platform power
There’s no need to worry that AI will deceive people into making foolish decisions. However, researchers warn that AI will significantly reduce the cost of personalized persuasion. Previously, it was expensive to create tailored messages for large audiences. Now, AI can produce millions of unique messages for millions of people at almost zero cost. This type of “super persuasion” is beyond what previous communication technologies could achieve.
More dangerously, AI may fabricate false evidence to increase its persuasiveness. While previous AI errors were unintentional, it can now create fabricated research and examples to convince you. Therefore, the most practical strategy is to avoid relying on a single AI model for everything. Use different AI services for different tasks—use one to search for news, another to make purchases, and a third for professional information. Just as no one relied on a single website for all online content 20 years ago, using the same AI for everything exposes you to its overwhelming information flow, potentially leading you to biased conclusions without even realizing it.