Visibility into AI Agents [2024]

Описание: This paper explores the risks associated with increasingly autonomous AI agents and proposes measures to improve visibility into their deployment and operation. It argues that as AI agents become more capable, they may exacerbate existing societal risks and introduce new ones due to their potential to remove humans from the decision-making loop. The authors suggest three categories of measures to increase visibility: agent identifiers, which distinguish AI agents from human actions; real-time monitoring, which allows for immediate intervention on problematic behaviour; and activity logs, which record inputs and outputs for post-incident investigation. The paper also considers the implications of these measures for privacy and the concentration of power, suggesting that a balance must be struck between accountability and individual privacy.

Authors: Alan Chan, Carson Ezell, Max Kaufmann, Kevin Wei, Lewis Hammond, Herbie Bradley, Emma Bluemke, Nitarshan Rajkumar, David Krueger, Noam Kolt, Lennart Heim, Markus Anderljung

Published: 23 Jan 2024

URL: https://arxiv.org/abs/2401.13138