AI agents tap humans as offline sensors

The "Human API" framing is the sharpest critique of agentic AI to appear outside academic circles — it names a dynamic that's already happening but has no governance framework yet.

• –Agents hit a physical-world wall and route around it by querying humans in users' social networks — people who never consented to being part of any AI system
• –A documented example: an OpenClaw agent called over 80 restaurants to harvest ingredient data, turning workers into unpaid, unconsenting survey respondents
• –Liability-shifting is structural: developers design confirmation prompts specifically to offload accountability onto the human who clicks "approve"
• –The author proposes sensing budgets (rate-limiting human queries like API calls) and auditable logs — concrete, tractable governance levers
• –The bystander problem — people who never interacted with the agent but get modeled through a user's contact graph — has essentially no legal protection under current frameworks