The curious experiment of the social network populated entirely by chatbots

Born as tools to shorten distances, encourage the sharing of ideas and create virtual communities, social networks have however shown a dark side that has unfortunately consolidated over time, generating negative phenomena that are not always easy to manage. In a study still under review and entitled Can We Fix Social Media? Testing Prosocial Interventions using Generative Social Simulation conducted by researchers at the University of Amsterdam, professor of artificial intelligence and social media Petter Törnberg and researcher Maik Larooij experimented with an unusual scenario: activating a platform populated exclusively by bots with the aim of assessing whether a digital environment devoid of human users could reduce the tendency of social networks to feed polarising phenomena.

More and more often, these platforms are perceived as spaces where haters, hate language and unethical behaviour proliferate: the relative anonymity and physical distance offered by the screen obviously contribute to lowering inhibitions, and what would rarely be said by looking a person in the eye, instead finds free rein through a comment typed in a few seconds from a smartphone; insults, threats and real campaigns ensue, which, in the most extreme cases, even lead to cyberbullying.

Another determining factor is added to this scenario: the functioning of the algorithms, which is not always perfect. Polarising content, capable of stirring up conflict, generates more interactions and greater visibility and, in this way, the logic of 'like' and sharing tends to reward confrontation over constructive confrontation, reinforcing the so-called echo chambers, i.e. digital environments in which opinions, albeit wrong, are not questioned but continuously confirmed.

But the real worrying fact highlighted by the study is another: it is not only humans but also bots that create these disturbances. During the experiment, in fact, the researchers set up a social network without advertisements, suggested posts and an algorithm deciding what to show the users, i.e. 500 chatbots based on GPT-4o mini OpenAI each with a different personality and certain political tendencies.

In five different experiments each with over 10,000 simulated actions, the team let the bots behave like normal users capable of writing posts, following profiles and re-sharing content. However, the result was surprising and reminiscent of what we see every day on real platforms: chatbots, in fact, tended to get closer to those who held the same opinions as them, forming 'bubbles' in which the same ideas were reinforced and more extreme posts received more attention and attracted more interactions than moderate ones. Over time, moreover, the conversation became concentrated in the hands of a few more active bots, which ended up dominating the scene, reproducing the mechanism reminiscent of real-life social networks, where a small number of influencers move much of the debate by conditioning the visibility of other content.