Joelle Renstrom is a science writer and teaches at Boston University.
Trying to change someone’s mind in an online or social media exchange is about as effective as flapping your arms to fly. That’s why a recently published study in Nature Human Behavior might seem surprising: It found that AI can succeed in this apparently impossible task. In fact, AI can be considerably more persuasive than humans in online debates. The reasons provide insights into the shortfalls of human interactions and how we can better communicate with one another.
The study measured the effectiveness of ChatGPT-4 versus humans at changing minds. It found that when neither the human nor the AI knew anything about their interlocutor, they were equally persuasive. However, when ChatGPT had basic demographic information about the person, such as age, ethnicity, and sex, it outperformed its human counterpart 64 percent of the time.
The study’s finding is consistent with previous studies that have found AI to be more convincing than humans in dissuading people from believing conspiracy theories. The question is: Why?
AI tailors its arguments differently when it knows something about a user than when it does not. With background info, AI can identify common ground before making its case. If AI knows, for example, that it’s dealing with someone with a liberal political affiliation, it will use arguments popular among liberals, such as the importance of immigrants to the national culture and economy. If a user leans conservative, the AI highlights ideals such as lack of government intervention or regulation. That means an AI with even fairly basic information about users can make them more receptive by tapping into their confirmation bias — in other words, by affirming their existing beliefs.
This might seem manipulative, but the AI is assessing the rhetorical situation, which involves figuring out what will and won’t appeal to a specific audience and adjusting arguments accordingly. I teach this concept to my rhetoric students at Boston University. When presented with evidence that runs counter to their views, many people cling even more firmly to their positions — a phenomenon known as the “backfire effect” or “belief perseverance.” But much of the time, appealing to an audience starts with sharing common ground.
Of course, humans should be better than AI is at understanding and bonding with fellow humans. Why aren’t we? For one, our ability to “read the room” has declined as interactions with devices have skyrocketed. Even online, it’s easy to glean basic facts about fellow human users from photos and posts, so that we can adjust our talking points accordingly. But all too often, we don’t. What’s more, online algorithms ensure that users stay in echo chambers that reinforce their own points of view. When humans venture out of those bubbles, they often do so spoiling for a fight. That’s not helpful for changing minds. AI succeeds in part because it does the opposite.
AI doesn’t shout, curse, judge, threaten, or doxx users. That’s one reason AI can more easily change someone’s mind: Freedom from judgment and retaliation liberates the mind enough to allow people to consider other viewpoints.
AI also deploys an age-old human bonding technique: mirroring. Experiments have demonstrated that if an interactive robot smiles at you, you’ll probably smile back, and vice versa. We all do this beginning in infancy as a way of expressing and emulating other people’s behavior and emotional states. If an AI discerns that you’re smiling, it might tell you a joke so it can capitalize on and match your positive mood. Mirroring can also involve language choice and tone, which is another reason AI excels at it. Establishing even a fleeting bond via mirroring makes humans more likely to cooperate with an AI or robot.
Perhaps AI succeeds in part because humans so often fail at effective communication. The bar has become frighteningly low for interactions, both online and off. If displaying a modicum of respect and not spewing vitriol gives AI a communication advantage, that’s on us. If a noncorporeal, nonhuman entity can read the room and connect with people better than most humans can, something is seriously off.
Research into AI reveals what’s too often missing in human interactions: common ground, curiosity, and compassion. Perhaps two of the ways AI will serve us best are by reflecting the failures of human societies in its black mirror and reminding us that we can do better. In turn, that perspective can help us choose what or who we want to become — and how.