AI is reshaping the dynamics of persuasion in ways that are likely to affect public safety, democratic processes and social cohesion. Joanne Hinds, Dessi Bocheva and Patricia Hajszan at the University of Bath examine the ability of AI systems to generate persuasive content at scale and the complex challenges this presents.
Recent advances in artificial intelligence (AI) are changing how persuasion operates in digital environments. Large language models (LLMs), such as those used in tools like ChatGPT, can generate coherent arguments, respond to users in real time and adapt to conversational cues. As a result, persuasion is no longer limited to static messages. It can now take the form of ongoing, interactive exchanges that respond to individuals’ beliefs, values and motivations.
For policymakers, this raises questions about how public opinion is formed and how it may be shaped over time through everyday digital interactions. The primary risk may not lie in the persuasiveness of individual messages, but in the scale, speed and coordination with which AI-generated content can be deployed.
From static messages to interactive persuasion
The ability to persuade, and our propensity to be persuaded, are fundamental to how modern societies function. From marketing campaigns that shape consumer preferences and drive purchasing decisions, to public health messages that promote healthier lifestyles, persuasion plays a central role in guiding what we think and do. Even digital platforms have been designed to nudge where we click, what we engage with and how long we stay online.
Today, a growing body of research suggests that AI systems influence attitudes and judgements across a range of topics. It has been found that AI-generated arguments can be as persuasive as those written by humans, potentially altering people’s views on public policy issues such as carbon tax and gun control. Interaction with AI systems can also shift people’s views and their confidence in those views, even when persuasion is not the system’s explicit aim.
Importantly, these effects may not be short lived. Evidence suggests that changes in attitude following interaction with AI can persist for weeks. While much of this research comes from controlled experimental settings, it indicates that conversations with AI systems have the potential to shape beliefs beyond the immediate interaction.
From individual persuasion to information environments
AI is already being integrated into real-world influence operations. Transparency reports from AI developers document cases where malicious actors have attempted to use AI to generate political content at scale. In one instance, accounts linked to a China-based operation used AI tools to produce polarising social media posts on both sides of the political spectrum. Other cases show AI being used to analyse online discourse, generate spam and scam material, and distribute it widely.
These cases show how AI can amplify persuasive influence at scale. Whereas traditional influence campaigns required significant human effort to craft messages, distribute content and maintain online personas, AI systems can automate much of this process. Large volumes of persuasive content can be produced in seconds and at a low cost. When deployed across networks of accounts, this capability can reinforce particular narratives, create the appearance of widespread agreement and shape the wider information environment in which people form their views.
Detection challenges
AI-generated content may also be particularly convincing because of its structure. Research suggests that such content is often more grammatically consistent, information dense and logically structured than human writing. These features can act as cues of credibility: people tend to interpret fluency, coherence and detail as signals of expertise. This leads people to perceive the content as highly convincing, even when the underlying information is flawed or misleading.
Studies also show that people struggle to distinguish between human-written and AI-generated content, including in contexts such as product reviews and phishing messages. Notably, greater familiarity with AI systems does not consistently improve people’s ability to detect such content. This creates a challenging dynamic: even when people are aware that AI-generated content exists, they may still rely on superficial cues, such as fluency or confidence, when deciding what to believe. For this reason, recognition alone may not be enough to guard against the influence of AI-generated content, a pattern sometimes described as a recognition-behaviour gap.
A dual-use technology
The same capabilities that enable AI-generated content to spread misleading information can also be used for public benefit. Emerging evidence suggests that conversational AI can help to counter misinformation and reduce conspiracy beliefs. Some interventions lead to sustained reductions in false beliefs over time. More broadly, AI-generated explanations can improve people’s ability to identify misleading information online.
This highlights that AI-driven persuasion is not inherently harmful. AI systems can be used to support informed decision-making and provide accurate explanations. In this sense, AI persuasion is best understood as a dual-use capability, with its impact depending on how these systems are designed, deployed and governed.
Adapting to AI-driven persuasion
As AI becomes embedded in everyday digital interactions, the question is not whether it can persuade, but how it changes the broader information environment and how societies can adapt. Persuasive messages can now be generated instantly, tailored to individuals and distributed at unprecedented scale. This has the potential to shape public attitudes, undermine trust in information sources and influence how people engage with democratic processes.
For public safety and security policymakers, the priority is to understand how AI systems reshape the information environment. Even if the persuasive power of individual AI-generated messages is modest, the ability to produce and distribute persuasive content in large volumes makes it more difficult to maintain information integrity online. It also makes it more difficult to identify and respond to coordinated activity. This suggests a need to strengthen approaches to monitoring and responding to influential messaging, while improving public resilience to misleading information. Addressing these challenges will require not only technical solutions, but also careful decisions about how AI systems are designed, governed and used in practice to support a more informed, trustworthy and resilient information environment.
All articles posted on this blog give the views of the author(s), and not the position of the IPR, nor of the University of Bath.
