The dominant conversation around AI safety tends to fixate on dramatic, world-ending scenarios: an AI-enabled nuclear war, a deliberately engineered biological outbreak. A new arXiv preprint argues that this narrow focus may be obscuring a quieter but equally important class of danger. The paper, a position statement rather than an empirical study, contends that AI risks can also be accumulative — emerging not from a single catastrophic event but from many small, compounding effects over time.
To make this case, the authors turn to social media. Decades of platform deployment have shown how seemingly minor design choices can gradually reshape public discourse, mental health, and political polarization. These harms were not the result of one decisive failure, but of millions of incremental interactions. The paper suggests that AI systems, which are already embedded in similar feedback loops, may pose analogous risks that are poorly captured by existing safety frameworks.
The authors do not dismiss existential threats, but they argue that a singular focus on them leaves a blind spot. If accumulative risks are real, then safety evaluations need to account for long-term, systemic effects — not just worst-case outputs. The preprint is a call to broaden the research agenda, using social media as both a warning and a template for studying how AI can harm us slowly rather than all at once.