Social platforms may be wired to amplify polarization by their structure alone. That is the suggestion of a 2025 study by Maik Larooij and Petter Törnberg, shared as a preprint before peer review. Their paper, “Can We Fix Social Media?”, describes a simulated platform in which every user was an AI agent, a program that acts on its own. The simulation reproduced three familiar failures. Users sorted into echo chambers that mostly showed them views like their own, influence gathered in a few accounts, and polarization rose.
Agents in the model did only three things: they posted, followed one another and reposted. No extremists were seeded, and no toxic intent was built in. Still, the network drifted into the same distortions seen online.
Next, the researchers tested six common-sense fixes. They showed posts in time order, hid reposts and follower counts, and boosted diverse viewpoints. They also gave priority to empathetic content, limited the dominance of high-reach accounts and removed bios and identity cues. None of the fixes changed the results much, and some made outcomes worse.
The same drift can happen on a public platform, in a brand community or in an internal company network. Each runs on default incentives of speed, engagement and reach, so tweaking algorithms or moderating bad actors leaves the deeper mechanics in place. Polarization is built into how influence and visibility accumulate.
For a business running any of these networks, that is a serious economic risk. Systems that chase clicks and reach build up long-term costs in lost trust, hostile communities, heavier moderation, brand damage and users who leave. Once polarization sets in, fixing it gets more expensive.
The study also shows how generative AI can simulate a platform’s behavior before launch. The researchers built the agents on AI language models, which let them test structural redesigns without needing a live user base. A company running a platform or community could test its feedback loops and alternative designs the same way before rollout. The authors flag one limit. A simulation only reflects its inputs, and the language models behind the agents were trained on biased, echo-prone data.
