Conversations with AI chatbots can significantly influence the direction of the vote

Both studies are very robust and, moreover, agree in their main conclusions, although, of course, they are not without limitations. The study by Lin et al. has a pre-registered design, involved citizens from several countries (USA, Canada, and Poland), and offers a rigorous analysis that reinforces the plausibility of its results. Meanwhile, the study by Hackenburg et al. stands out for its scale, both in the number of participants (almost 77,000) and conversations (91,000), as well as the linguistic models involved (both open and commercial weights), which allowed for a systematic evaluation of the various technical aspects behind the persuasive capacity of chatbots. In both cases, despite the artificiality of the experimental environments, these are exhaustive, transparent, and very interesting studies.

Both articles concur with the available evidence highlighting the power of dialogue and reasoning based on facts and evidence as the primary means of persuasion, but they also offer interesting new insights. For example, although most of the available literature states that persuasive messages should be personalized according to the audience's values, researchers have not found that such personalization offers significant benefits. They show how conversations with chatbots generate persuasive effects superior to those observed with traditional political messages, and, moreover, the main mechanism of persuasion is not psychological strategies, but rather the presentation of a high density of factual and verifiable information.

However, there is an important trade-off: when the chatbot is asked to be more persuasive, there is a risk that the amount of inaccurate information it offers will be greater. Interestingly, the articles also show a systematic asymmetry in the factual accuracy of the information generated by chatbots: when they have to persuade people to vote for right-wing candidates, they offer more inaccurate information than when they have to persuade people to vote for left-wing candidates.

The main implication is that the scenario they describe has ceased to be hypothetical and has become possible (and worrying): using chatbots to persuade citizens to vote a certain way through a dialogue supposedly based on facts, but which, in light of its results, may sacrifice factual accuracy in order to increase persuasive power. However, it's important to note that neither article claims that inaccurate information is more persuasive, only that as persuasive power increases, the amount of inaccurate information tends to increase.

[Regarding possible limitations] Both articles describe experimental situations with characteristics that may lead to different results in real-world contexts and campaigns. Thus, in the work by Lin et al., we must consider that:

1. Participants enrolled voluntarily (self-selection bias).
2. A controlled dialogue (where the human knows they are participating in an experiment involving a chatbot) is not the same as a campaign situation.
3. Real-world behavior, such as voting, was not measured.
4. The effects in the US were less pronounced than those observed in other countries.
5. The fact that chatbot persuasion can lead to the dissemination of inaccurate information opens the door to a series of significant risks (and ethical dilemmas) that require further attention.

Regarding Hackenburg et al., this study also has limitations:

1. It was conducted only in the UK and, consequently, cannot be directly generalized to any other country.
2. The participants were paid, which further distances the experimental conditions from real-world scenarios.
3. The requirement for informed consent and a prior briefing also distances the experimental conditions from those of a real campaign.
4. Again, the problem of the imbalance between persuasion and truthfulness remains unsatisfactorily resolved.

These studies are methodologically strong and highly relevant. They rely on very large samples, carefully designed experiments, and transparent outcome measures of opinion change following interaction with AI systems. The central result — that short conversations with large language models can produce measurable shifts in political attitudes — is robust across different contexts and datasets. The accompanying commentary in Science correctly emphasizes that these systems are not “superhuman persuaders” in a psychological sense, but are effective because they systematically generate dense streams of information, regardless of its truthfulness.

What these papers truly contribute is not simply the demonstration that AI can persuade, but an explanation of why it does so. The evidence shows that persuasion increases primarily through information density, not through personalization, emotional manipulation, or ideological targeting. The studies also show that post-training aimed at persuasiveness is substantially more important than model size. Most importantly, there is a clear and troubling trade-off: the same techniques that maximize persuasive impact systematically reduce factual accuracy. Persuasiveness and truthfulness do not grow together; they diverge. This means that the mechanism driving influence is not understanding, but volume and fluency.

This is where the deeper significance of these results becomes visible. These findings point toward what I call an “epistemic shift,” or Epistemia: a transformation in how knowledge operates in the public sphere. For decades, digital platforms primarily mediated information through filtering and ranking. Generative systems do something fundamentally different: they replace information retrieval with language synthesis. In doing so, they bypass the cognitive processes that normally structure judgment, verification, and evaluation. We are not dealing with machines that lie. We are dealing with systems that generate plausible language without performing any epistemic act at all.

The danger, therefore, is not only misinformation. It is something more structural. When information is generated rather than assessed, plausibility replaces judgment. These experiments show this displacement very clearly: participants are persuaded not by the quality of arguments, but by their quantity. Whether statements are true or false becomes secondary to the sheer accumulation of claims.

There are also important limits to underline. These experiments measure short-term shifts after brief interactions, whereas real-world exposure is continuous, immersive, and cumulative. For this reason, the measured effects should not be interpreted as upper bounds. If anything, they are likely conservative. Moreover, participants in these studies were aware that they were interacting with an AI. In everyday environments, where generative systems are integrated into search engines, messaging platforms, and productivity tools, contextual trust may amplify the effects further.

In summary, the core risk highlighted by this work is not merely that AI can influence opinions. It is that AI normalizes an informational environment where judgment is replaced by generation, and evaluation is replaced by fluency. This is not just a technological problem. It is an epistemic one — and these studies are among the first to demonstrate it empirically.