Social media toxicity can’t be fixed by changing the algorithms

The polarizing impact of social media is not only the result of bad algorithms – it is inevitable because of the fundamental components of the functioning of platforms, revealed a study with users generated by AI. He suggests that the problem will only be resolved if we fundamentally reinvent the world of online communication.

Petter Törnberg at the University of Amsterdam in the Netherlands and his colleagues created 500 AI chatbots designed to imitate a range of political convictions in the United States, based on the American National Election Studies Survey. These robots, fed by the GPT-4O mini-language model, were then responsible for interacting with each other on a simple social network that researchers had designed without ads or algorithms.

During five points of the experience, each involving 10,000 actions, AI agents tended to follow the people with whom they shared political affiliations, while those who have more partisan opinions have won more followers and republiers. This has echoed the overall attention to these users, who have gravity to more partisan posters.

In a previous study, Törnberg and his colleagues explored if the social networks simulated with different algorithms could identify the ways to write political polarization – but the new research seems to contradict their previous results.

“We expected that [polarisation] Being something that is motivated by algorithms, ”says Törnberg.”[We thought] That platforms are designed for this – to produce these results – because they are designed to maximize engagement and to piss you off and so on. »»

Instead, they found that it was not the algorithms themselves that seemed to cause the problem, which could make attempts to eliminate the antagonistic behavior of users by very difficult design. “We have set up the simplest platform we can imagine, then, Boom, we already have these results,” he said. “It already suggests that it stems from something very fundamental for the fact that we have a publication, republication and following behavior.”

To see if these behaviors could be silent or contracted, the researchers have also tested six potential solutions, including a food only chronological, by granting less prominence to viral content, by amplifying opposite views and in empathic and reasoned content, by hiding a disciple and republication of republication and by masking profile bios.

Most of the interventions have made little difference: the inter -party mixture has not changed more than 6%, and the share of attention is mounted by the most important accounts between 2 and 6% – while others, such as the hidden of the biographies of the involved users, actually aggravated the problem. When there have been gains in an area, they were countered by negative impacts elsewhere. The fixes that have reduced user inequality made extreme messages more popular, while changes to softening the partisanry have even more attracted attention to a small elite.

“Most social media activities are always fruits of the toxic tree – the starting problems of social media always reside in their fundamental conception and can as incorporation the worst human behavior,” explains Jess Maddox at the University of Georgia.

While Törnberg recognizes that experience is a simulation that could simplify certain mechanisms, he thinks he can tell us what social platforms must do to reduce polarization. “We may need more fundamental interventions and need to rethink more fundamental,” he says. “It may not be enough to stir with algorithms and modify the parameters of the platform, but [we might] Need to rethink the structure of interaction more fundamentally and how these spaces structure our policy. »»