What Social Media Teaches Us About Accumulative AI Risks

Accumulative harms deserve equal attention to catastrophic risks: While AI safety discussions often focus on dramatic scenarios like AI-enabled nuclear conflict, the paper argues that everyday harms accumulating through repeated AI use—such as gradually narrowed perspectives or weakened independent judgment—pose significant risks that are currently overlooked.

Social media provides a cautionary blueprint for AI deployment: The paper demonstrates that mechanisms observed in social media (reciprocal effects, social reinforcement, and selective exposure) can create harmful feedback loops without any single catastrophic event, suggesting similar patterns may emerge as AI systems become more integrated into daily life.

Feedback loops in AI systems can enable subtle exploitation: Using conversational AI as an example, the paper identifies how persistent memory, repeated interactions, and semantic inference can tighten feedback loops that enable sycophancy, manipulation, and privacy loss—harms that accumulate gradually rather than occurring as discrete incidents.

Current safety assessments miss "less visible harms": The Trajectory Neglect Thesis argues that focusing on identifiable incidents and vivid catastrophic events causes safety evaluations to miss slower-developing harms that emerge through widespread, interconnected AI deployments across society.

Mechanisms of harm require investigation beyond individual encounters: Rather than assuming AI risks operate the same way as isolated incidents, the paper calls for research into how individual susceptibility, social reinforcement, and system design interact to create accumulating harms over time.