Compiled by the editorial desk with reference to the University of Amsterdam preprint and reporting by Ars Technica.

Researchers at the University of Amsterdam have built a social network populated entirely by AI chatbots to test whether common design changes can reduce political polarization. The results, detailed in a preprint not yet peer-reviewed, suggest that even well-intentioned interventions often fail—and occasionally make things worse.

The study, led by Petter Törnberg, an assistant professor specializing in AI and social media, and research assistant Maik Larooij, simulated a platform where each user was driven by OpenAI's GPT-4o large language model. They tested six interventions, including switching to chronological feeds, promoting diverse viewpoints, hiding follower counts, and removing account bios. Their goal was to see if any could prevent the formation of echo chambers and toxic discourse.

According to the findings, none of the strategies achieved a satisfactory reduction in polarization. Some showed only modest improvements, while others backfired. For instance, a chronological news feed reduced attention inequality but pushed extreme content to the top, as reported by Ars Technica.

Why Interventions Fall Short

Törnberg explained that polarization is not merely a result of provocative content. "It also shapes the network structures that are formed," he told Ars Technica, "which in turn feeds back what content you see, resulting in a toxic network." This creates an "extreme inequality of attention," where a tiny minority of posts dominate visibility.

The findings challenge the vision, once articulated by Elon Musk, of social media as a "digital town square" where diverse voices coexist. Instead, the study suggests that without structural changes, platforms may remain breeding grounds for extremist thinking.

The researchers acknowledged that AI simulations are not perfect. Törnberg admitted that using AI is not a "perfect solution" due to "all kind of biases and limitations," but he argued that the technology can capture "human behavior in a more plausible way" than traditional models.

Implications for the Age of Generative AI

The study arrives as generative AI is already being used to produce content on platforms like X. Törnberg noted that many actors are "using AI to produce content that just seeks to maximize attention," often spreading misinformation and highly polarized information. As AI models become more powerful, he warned, "that content is going to take over."

He expressed doubt about the survival of conventional social media models in such an environment. "I have a hard time seeing the conventional social media models surviving that," he said.

The findings underscore the difficulty of designing online spaces that foster constructive debate, a goal that has eluded platforms since the early days of the internet. While the study is preliminary, it offers a sobering perspective for policymakers and tech companies seeking to address online toxicity.