Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Role of Governments in Increasing Interconnected Post-Deployment Monitoring of AI

Merlin Stein, Jamie Bernardi, Connor Dunlop

TL;DR

This work highlights example data sources and specific data points that governments could collect to inform AI risk management and suggests inference time monitoring of chain-of-thought reasoning can be combined with long-term monitoring of sectoral AI diffusion, impacts and incidents.

Abstract

Language-based AI systems are diffusing into society, bringing positive and negative impacts. Mitigating negative impacts depends on accurate impact assessments, drawn from an empirical evidence base that makes causal connections between AI usage and impacts. Interconnected post-deployment monitoring combines information about model integration and use, application use, and incidents and impacts. For example, inference time monitoring of chain-of-thought reasoning can be combined with long-term monitoring of sectoral AI diffusion, impacts and incidents. Drawing on information sharing mechanisms in other industries, we highlight example data sources and specific data points that governments could collect to inform AI risk management.

The Role of Governments in Increasing Interconnected Post-Deployment Monitoring of AI

TL;DR

This work highlights example data sources and specific data points that governments could collect to inform AI risk management and suggests inference time monitoring of chain-of-thought reasoning can be combined with long-term monitoring of sectoral AI diffusion, impacts and incidents.

Abstract

Language-based AI systems are diffusing into society, bringing positive and negative impacts. Mitigating negative impacts depends on accurate impact assessments, drawn from an empirical evidence base that makes causal connections between AI usage and impacts. Interconnected post-deployment monitoring combines information about model integration and use, application use, and incidents and impacts. For example, inference time monitoring of chain-of-thought reasoning can be combined with long-term monitoring of sectoral AI diffusion, impacts and incidents. Drawing on information sharing mechanisms in other industries, we highlight example data sources and specific data points that governments could collect to inform AI risk management.
Paper Structure (11 sections, 2 figures, 1 table)

This paper contains 11 sections, 2 figures, 1 table.

Table of Contents

  1. Interconnected Post-Deployment Monitoring of AI as a Government Priority
  2. What is Post-Deployment Monitoring of AI Systems?
  3. Types of Post-Deployment Information
  4. Deployment configurations: Different Supply Chain Actors’ Possession of Information
  5. Challenges for Governments conducting Post-Deployment Monitoring of AI
  6. Recommendations for Governments on Post-Deployment Monitoring
  7. Prioritise Incident Monitoring and Reporting with causal links to AI system use
  8. Establish Mechanisms to Gather Post-Deployment Information
  9. Request Initial Data Points and Build Analysis Capacity
  10. Support Technical Governance Methods that Increase Visibility
  11. Conclusion and Future Work

Figures (2)

  • Figure 1: Information types for AI governance, categorised by supply chain actors. Some information sharing involves structured documentation (disclosure-focused), some requires additional analysis (assessment-focused). From Stein and Dunlop steinDunlopAda2024. Information subcategories are superscript, and are drawn from 1) the Foundation Model Transparency Index bommasani2024foundation, 2) the International Scientific report on the Safety of Advanced AI internationalTechnicalReport and 3) the Sociotechnical Safety Evaluation Repository weidinger2023sociotechnicalsafetyevaluationgenerative.
  • Figure 2: Example deployment configurations including some notable AI systems. On the left, we present the foundation model supply chain from JonesExplainer2023. On the right are five example deployment configurations (non-exhaustive).