For years, digital politics has been selling the same promise: getting to know a voter better and better, collecting ever more precise data, tapping into their habits, their values, even their exposed nerves, to deliver the perfect argument. It’s the dream of microtargeting: the message tailored to the right person, at the right time, with the right tone. Then there is the other great article of faith, the more academic and less advertising one: the belief that a position really changes when the listener is dragged into a long, demanding reasoning, full of questions, objections, defenses, mental effort.
A study published in PNASone of the most cited science journals in the world, took these two ideas and brought them to the test bed with artificial intelligence. The result has a clear ring to it: chatbots can shift political opinions, but personalized messages and deep conversations don’t seem to beat a simple, well-written generic topic.
The right message
Political persuasion has always weighed heavily, and it weighs a lot. Pressure groups, candidates, health institutions, committees, foundations: they all spend money and time to convince citizens who are often already polarized. The problem comes when you try to understand precisely why a person changes their mind. In laboratories, real communication comes in through the door crookedly. A real researcher changes tone, an extra involuntarily smiles, a pause sounds like social pressure, a look alters the meaning of a sentence. The participant’s mind records everything, even what the experimental protocol would have liked to keep out.
This is where the study comes in. The starting idea was almost brutal in its simplicity: to use large linguistic models as controllable, constant, large-scale replicable debate partners. Same style, same discursive posture, same structure of the topic, with the possibility of varying only one element at a time and seeing what really happens when personalization comes into play or when one tries to increase cognitive processing, i.e. that mental effort which according to the elaboration likelihood model should leave more stable traces in attitudes.
To do so, the team built two pre-registered online experiments with nearly 3,700 U.S. adults, recruited to approximate census averages for age, gender and race, with a political balance sought from the start between Democrats and Republicans. The first study was about immigration: more funding for border security or more openness to sponsored immigrant visas. The second entered into another incandescent terrain, that of school curricula: how much power parents should have over controversial social issues addressed at school and how far teachers can go with their political opinions in the classroom. Two fields well chosen, because in the United States a few minutes on these issues are enough to understand how the public debate has become rigid, identitarian, often tired of listening.
After recording initial opinions, the researchers divided the participants between the control group and four interventions based on a linguistic model. In any case, the bot had a very specific task: to support the opposite thesis to that expressed by the participant. The first group received a single generic text, written as the best possible paragraph supporting the opposing position. The second received a microtargeted message, constructed using the demographic data provided at the beginning of the survey. The third entered into a six-round direct confrontation with the AI, instructed to behave like a psychology expert capable of retorting and asking questions to increase mental involvement. The fourth group took part in a kind of motivational interview, a technique often used in the therapeutic field, in which the bot tried to push the participant to find the reasons for the change on their own.
Where microtargeting deflates and only the strength of a well-written argument remains
To avoid the most obvious objection, the researchers also checked that the underlying contents remained comparable. With machine learning tools they mapped the argumentative cores of the messages produced by the system and verified that the most visible differences were in the packaging, in the way of presenting the material, in the type of interaction, while the factual heart of the arguments remained substantially aligned. In other words, the studio sought to separate dress from substance. And that’s exactly where the dress stopped seeming miraculous.
There was a persuasive effect. This must be said clearly. Exposure to the opposing argument pushed many people to moderate their position, with an average shift estimated between approximately 2.5 and 4 percentage points in the direction of the argument received. The interesting fact comes a second later: the more sophisticated methods delivered little more than the basic message, and often didn’t yield more at all. Deep personalization and interactive chats have not shown a compelling advantage over a single generic paragraph. In the immigration experiment, motivational interviewing was even among the least effective approaches. For those who imagine electoral campaigns dominated by machines capable of reading the soul of the individual voter, the blow is remarkable.
This passage puts two very deep-rooted ideas into difficulty. The first concerns microtargeting, presented for years as the decisive lever of digital political communication. The second touches on the elaboration likelihood model, that is, the belief that the most solid change comes when the person has to commit, respond, reason and defend their position. The study does not say that these mechanisms do not exist or that they are worth zero. He says something more annoying, and for this reason more useful: within a short and controlled interaction, the additional advantage seems small, much smaller than the rhetoric of campaigns and consultants promises. Sometimes a linear, well-constructed argument, expressed coherently, is enough to obtain almost the same effect as the refined and expensive device.
Opinions move, hostility remains steadfast
The researchers also looked elsewhere, and here the picture becomes even more interesting. In addition to changing positions on individual policies, they measured democratic reciprocity, that is, the willingness to consider political opponents reasonable people, worthy of respect, legitimate within the democratic space. For years, many scholars have wondered whether reducing the distance on an issue also means lowering hostility towards the group that thinks differently. That would be reassuring. It would also be very convenient. The data collected here tells a rougher scene.
People, in several cases, have tempered their policy opinions. The feeling towards the opposing camp, however, remained almost where it was. The ideological moat has narrowed a little, the underlying antipathy towards the other political bloc has held its ground. There was only one notable exception: in interactive chats about school curricula, participants showed an increase in democratic reciprocity. The authors hypothesize that it depended on the fact that, in that specific context, the bot openly insisted on the value of social tolerance within educational discourse. It’s a tiny detail, and for this reason it’s worth a lot: shifting an opinion doesn’t automatically open up respect, trust, mutual recognition. Politics remains full of people who can approach each other on a concrete level and continue to look at each other with the same harshness as before.
The real lesson also concerns social research
The authors urge not to take these results as a definitive ruling. The experiments observe short interactions, which occur in an isolated digital environment. Real life works with longer times, stratified relationships, shared memories, reputations, faces, silences, embarrassments, social obligations. A topic uttered by a close friend, a family member, an esteemed person, enters the mind differently than the same topic read in an online survey. Here AI doesn’t erase the problem, it puts it in a cleaner light. The main promise of the work, in fact, also lies in the method: using generative systems to produce thousands of controlled, comparable, repeatable interactions, with much lower costs than an entirely human experimental apparatus.
And this is perhaps the part that will remain the longest. Artificial intelligence, in this case, does not appear as an oracle or as a threat from a dystopian film. It appears as a laboratory tool capable of putting pressure on theories that seemed established. The work published in PNAS shows that political persuasion via AI works, and this alone is enough to raise the antennas. It also shows that the fantasy of the surgical message, hyper-personalized, capable of digging much deeper, after all, holds up less than expected when you look at it closely. For a sector that has been selling emotional precision for years as if it were an exact science, the news is significant. Much more than a well-written paragraph.
You might also be interested in:
