Mapping the individual, social, and biospheric impacts of Foundation Models

Current AI governance is missing urgent, real-world harms: Major international AI governance initiatives like the UK's AI Safety Summit and G7's Hiroshima process have over-emphasized technical safety and existential risks while neglecting documented harms already occurring, including discriminatory outputs, privacy violations, copyright infringement, and exploitative labor practices.

Foundation models create widespread harms across three interconnected levels: The paper identifies 14 categories of risks affecting individuals (such as discriminatory or toxic outputs), society (including labor exploitation and copyright violations), and the environment (through resource extraction and energy consumption), requiring governance approaches that address all three dimensions rather than focusing narrowly on one.

Environmental and labor costs are being systematically overlooked: The research highlights that the "brute-force industrialization" of AI applications involves unjust extraction of both natural resources and labor, impacts that remain largely invisible in mainstream AI governance discussions despite their significant real-world consequences.

A broader framework is needed for responsible AI policy: Rather than focusing primarily on speculative catastrophic risks, the paper argues that AI governance must integrate social, political, and environmental considerations to address the full spectrum of harms that foundation models and generative AI are already producing.