Visibility into AI Agents

Increased delegation of commercial, scientific, governmental, and personalactivities to AI agents – systems capable of pursuing complex goals withlimited supervision – may exacerbate existing societal risks and introduce newrisks. Understanding and mitigating these risks involves critically evaluatingexisting governance structures, revising and adapting these structures whereneeded, and ensuring accountability of key stakeholders. Information aboutwhere, why, how, and by whom certain AI agents are used, which we refer to asvisibility, is critical to these objectives. In this paper, we assess threecategories of measures to increase visibility into AI agents: agentidentifiers, real-time monitoring, and activity logging. For each, we outlinepotential implementations that vary in intrusiveness and informativeness. Weanalyze how the measures apply across a spectrum of centralized throughdecentralized deployment contexts, accounting for various actors in the supplychain including hardware and software service providers. Finally, we discussthe implications of our measures for privacy and concentration of power.Further work into understanding the measures and mitigating their negativeimpacts can help to build a foundation for the governance of AI agents.

Further reading#

• Access Paper in arXiv.org