Submission to the Global Digital Compact on the Interpretability of AI Foundation Models

· 3 min read
By , and

As part of the UN’s Our Common Agenda, an intergovernmental process on digital governance has been put in place to produce the ‘Global Digital Compact’. The Office of the Tech Envoy and the co-facilitators, Rwanda and Sweden, are now leading a consultative period to outline an international tech governance roadmap.

The GDC represents a significant step forward in building multilateral capacity to deal with some of the great technological questions of our age, including the rapid rise of artificial intelligence. The Simon Institute consulted various stakeholders from the private sector, academia, and civil society on AI safety and governance. SI also participated in the Future of Life’s consultation process; consulted with Chinese collaborators to ensure international coherence; and convened 3 workshops on the impacts and governance of AI R&D. Our submission to the GDC synthesizes all these efforts to engage diverse stakeholders on AI governance. 

Summary

In the last 6 months, the Simon Institute for Longterm Governance held three workshops on future-proofing the UN. Recommending a focus on AI governance to the Global Digital Compact was identified as important by policymakers and technical experts from 6 continents. After conducting an analysis of other submissions and consulting all key stakeholder groups, a contribution on the governance of foundation models was judged most valuable. These powerful AI systems cannot currently be reliably aligned with human values, unless increased investment in interpretability research makes these black-box algorithms more accessible.

Principles

  1. Emphasize the convergence of problems and solutions in AI regulation, and thus advocate for the development of reliable interpretability tools.
  2. Balance expertise with democratic input, engaging with private sector and academia while ensuring citizens have a say in designing the future.
  3. Advocate for accountability of foundation models, as current regulation is ambiguous regarding the liability for harms caused.
  4. Leverage AI for sustainable development while reducing low-probability high-impact risks from foundation models.

Recommendations

  1. UN AI capacity-building program: Offer scholarships for experts from low- and middle-income countries to improve AI expertise at leading AI centres, and hire experts to forecast AI capabilities, opportunities, and harms.
  2. International consensus on AI auditing: Advocate for international coordination of AI auditing organizations through a multilateral forum, similar to the International Atomic Energy Agency’s Regulatory Cooperation Forum.
  3. Prioritize international cooperation on sharing AI incidents: Promote public sharing of incidents, adopt interoperable standards for reporting AI incidents, and support whistleblowing rights for those who work on advanced AI systems.
  4. Interoperable standards for safe AI: Engage standard-setting bodies to promote international collaboration on safety that respect diverse regulatory frameworks while remaining interoperable. Private-sector innovation can also provide input.

You can read the submission in full here.