Study finds platform design, not algorithms, drives social media toxicity

A new study using AI chatbots to simulate social media interactions reveals that platform toxicity and political polarization aren’t primarily caused by algorithmic manipulation—they’re built into the fundamental structure of how social networks operate. The research suggests that efforts to reduce antagonistic behavior through algorithm tweaks alone are unlikely to succeed, requiring more radical reimagining of online communication platforms.

What you should know: Researchers at the University of Amsterdam created a controlled experiment using 500 AI chatbots with diverse political beliefs interacting on a simple social network with no ads or algorithms.

• The bots, powered by GPT-4o mini (an AI language model), were based on data from the American National Election Studies Survey to represent a realistic spectrum of US political views.
• Across five experimental runs involving 10,000 actions each, the AI agents consistently followed users who shared their political affiliations while more partisan voices gained larger followings and more reposts.
• Even without algorithmic manipulation designed to maximize engagement, polarization emerged naturally from basic social media mechanics.

The big picture: This research contradicts the widely held assumption that social media toxicity stems primarily from engagement-driven algorithms designed to keep users scrolling.

• “We were expecting this [polarisation] to be something that’s driven by algorithms,” says lead researcher Petter Törnberg at the University of Amsterdam. “[We thought] that the platforms are designed for this – to produce these outcomes – because they are designed to maximise engagement and to piss you off and so on.”
• Instead, polarization appeared to be an inevitable consequence of fundamental social media behaviors: posting, reposting, and following.

Testing potential solutions: The researchers evaluated six different interventions to combat polarization, but most proved ineffective or counterproductive.

• Interventions included chronological feeds, reduced prominence for viral content, amplifying opposing viewpoints, hiding follower counts, and concealing user profile information.
• Cross-party interaction improved by only 6 percent at most, while attention distribution among top accounts shifted by just 2-6 percent.
• Some fixes, like hiding user biographies, actually worsened polarization by removing context that might encourage more nuanced interactions.

Why traditional fixes fail: The study revealed a troubling pattern where improvements in one area created problems elsewhere.

• Solutions that reduced inequality among users paradoxically made extreme posts more popular.
• Interventions designed to soften partisan divisions ended up concentrating even more attention on a small elite group of users.
• “We set up the simplest platform we could imagine, and then, boom, we already have these outcomes,” Törnberg explains.

What experts think: The findings align with broader critiques of social media’s foundational design principles.

• “Most social media activities are always fruit of the poisonous tree – the beginning problems of social media always lie with their foundational design, and as such can encourage the worst of human behaviour,” says Jess Maddox at the University of Georgia.
• Törnberg acknowledges the simulation’s limitations but believes it points toward necessary structural changes: “We might need more fundamental interventions and need more fundamental rethinking.”

Looking ahead: The research suggests that meaningful progress requires abandoning incremental algorithm adjustments in favor of fundamentally restructuring how online social spaces operate.

• Rather than tweaking platform parameters, the solution may involve completely reimagining “the structure of interaction and how these spaces structure our politics.”
• This challenges the current approach of tech companies that typically focus on algorithmic modifications rather than architectural overhauls.