For years, influence campaigns followed a familiar, almost predictable playbook. Governments, political groups, and other actors tried to shape what people thought by flooding the airwaves with propaganda, seeding misinformation, and pushing messages designed to divide or persuade. That playbook is still very much in use. But now artificial intelligence has turned it into something faster, cheaper, and much harder to spot. Generative AI can manufacture enormous quantities of content in seconds, tailor that content to different audiences, and power automated accounts that chat, argue, and post in ways that feel remarkably human. The result is influence at a scale, speed, and level of coordination we have never seen before. In Venezuela, for example, pro-government messages were spread through videos starring AI-generated news anchors, giving the world what looked like independent international coverage when it was actually synthetic content designed to manufacture consent. But here is the thing: I believe much of the public conversation is focused on the wrong question. We keep asking, “How do we stop fake news?” or “Can we detect AI-generated content?” Those are important puzzles, but they miss the deeper issue. The real challenge is not simply whether people encounter misleading information. It is understanding how, when, and why online influence sometimes tips over into real-world action. That is one of the defining security challenges now facing modern democracies. Governments are caught between protecting free speech, legitimate political participation, and social cohesion, while also having to respond to foreign interference, coordinated manipulation, and politically motivated violence. And all of this is getting harder as trust in institutions falls, social media rewards outrage over nuance, and existing divisions become fertile ground for exploitation. Too often, our responses are reactive, coming only after damage has already been done. That is why my colleagues and I are researching a different approach, one that combines AI with decades of social and behavioural research to understand how influence moves through groups and, in some circumstances, becomes real-world mobilisation.
Let’s be honest with ourselves: influence is not inherently sinister. It is not some dark force that only bad actors use. Influence is the glue that holds human societies together. Every time we persuade a friend to see a movie, convince a neighbour to vote, or inspire a community to clean up a park, we are using influence. Political movements, advertising, activism, public debate, all of these depend on persuasion. AI has not rewritten the basic rules of influence. What it has changed is the scale, speed, and coordination with which influence can be applied. Technology has also given researchers access to enormous amounts of information. We can now look at millions of social media posts, track how language changes over time, and spot patterns that were previously invisible. But data alone cannot explain why an influence campaign succeeds with one group and completely fails with another. The reason is simple: people are not passive recipients of information. We do not just absorb messages like sponges. We interpret everything through the lens of our identities, our experiences, our relationships, and the groups we belong to. The exact same content can land very differently depending on who sees it, which group they feel attached to, and the social context in which it appears. This is why prediction must never be confused with explanation. AI can find patterns across millions of data points, but those patterns still need to be interpreted using genuine knowledge of human behaviour. That is where the work we are doing becomes so important. The strongest approach is to combine the sheer scale of data science with the theories that explain how people actually think, feel, and act. We cannot just throw AI at the problem and hope it gives us answers. We need to pair it with an understanding of the human heart.
One of the most important lessons from our research is that exposure is not the same as mobilisation. It is easy to look at the internet and think that everyone is being radicalised, that every polarised argument will eventually spill into the streets. But that is not what we see. Online influence is everywhere, yet offline action is comparatively rare. Polarisation does not necessarily lead to violence. Holding radical ideas does not necessarily mean someone will act on them. Influence and mobilisation are different things. The more important question is how people move from consuming or sharing content online to acting together offline. To answer that question, my colleagues and I study several connected but distinct processes: politicisation, polarisation, radicalisation, and activation. These words are often tossed around as if they all describe the same phenomenon, but they actually represent different pathways of political and social engagement. Understanding those pathways can help explain why influence campaigns so often fail, and why, in other circumstances, they contribute to serious real-world consequences. Our research shifts the focus away from the search for inherently “dangerous individuals” and toward the wider group dynamics that shape behaviour and societal outcomes. When we talk about declining trust, we are often describing one group losing trust in another. When we talk about polarisation, we are describing groups moving further apart in their views. Those divisions can then be identified and exploited by actors who want to deepen existing social fault lines. Looking for a lone wolf or a single evil mastermind is often too simple. The real story is usually about groups, relationships, and the emotional currents that run between them.
We should also remember that collective action is not something to be feared. Collective action occurs when people act as members of a group to advance that group’s position. It includes many forms of democratic participation, from peaceful protest and community advocacy to campaigning for political or social change. In Australia, outside of the three-year voting cycle, collective action is a key way that ordinary people interact with the political systems they are part of. Political engagement and protest should not, by themselves, be treated as threats. Peaceful collective action is essential to democratic life. Democracies depend on people being able to express grievances, advocate for change, and participate in public debate. Protest is like a valve. When that valve is closed off because conventional forms of activism become criminalised, people do not stop caring. If anything, our research from the lab shows that criminalising protest can actually be radicalising for some people. Denying people a legitimate outlet for their frustrations can push them toward more extreme ways of expressing themselves. The real challenge is identifying when authentic participation is being manipulated, artificially amplified, or channelled towards harm. That is a much more delicate task than simply shutting down protest or labelling all dissent as dangerous. We need to be able to tell the difference between genuine democratic engagement and coordinated manipulation. We cannot afford to treat every person who disagrees with the government as a threat, but we also cannot ignore the fact that some actors are trying to exploit those disagreements for their own purposes.
One strand of our research looks at what happens when opposing groups respond to each other over time. We studied the debate surrounding the Voice to Parliament in Australia, analysing social media posts published over fifteen months by prominent supporters and opponents. We used language-processing tools to examine anger, negativity, hate, and offensiveness, and then used statistical analysis to track whether changes in one group’s language predicted changes in the other group’s language the following week. The results were fascinating. Polarisation was not a simple or inevitable two-way spiral. Changes in hostile language among supporters predicted increased hostility among opponents in the following week, but the reverse pattern was not found. In other words, the response was uneven. One group appeared much more reactive to the language of the other. This matters because it shows that division does not grow uniformly across every group or situation. The pathway toward greater hostility is conditional. It depends on what is happening within and between specific groups at particular moments. Understanding which groups are reacting, what they are reacting to, and how those reactions change over time may provide a much more useful picture than simply measuring the total amount of hostile content online. It also tells us that intervention could be possible if we catch the right moment, before the pattern becomes entrenched. We cannot just look at the volume of anger; we have to look at the dynamics of the relationship.
This brings us to the next frontier of research. At Flinders University, we are exploring how AI, language models, and other computational tools can be combined with social psychology to map the role of emotions in confrontational action, identify language linked with growing commitment to political violence, and understand how language and movement networks shape one another. The aim is not to monitor opinions, label lawful dissent as dangerous, or suppress legitimate political expression. We have no interest in creating a surveillance state for thoughts. The aim is to understand changes in group behaviour, particularly the pathways and possible tipping points that may distinguish widespread online discussion from coordinated manipulation or a movement headed toward harm. Rather than only counting posts or identifying false claims, this work examines how narratives travel through groups, how people define who “we” are, how perceived threats are discussed, and how these patterns change over time. That kind of approach could give analysts a more informed basis for identifying emerging risks, while also reducing the danger of overreacting to genuine democratic participation. As information environments become more complex, democracies will need more sophisticated responses than simply removing content or debunking false claims after the event. The future of democratic resilience depends on understanding not just what information people see, but how identities, emotions, relationships, and group dynamics shape what happens next. AI is transforming the information landscape, but the human psychology that underpins influence remains remarkably consistent. Meeting the challenge therefore requires both strong methods that can detect patterns and strong theories that can explain them. If we want to protect trust, strengthen social cohesion, and safeguard democratic institutions, we need to understand technology and people together, not in separate boxes. We need to remember that behind every post, every share, every algorithm, there is a person trying to make sense of a complicated world. That is where the real work begins.