Drawing on deliberations at the recent AI for Good Summit in Geneva, Switzerland, Executive Director Pria Chetty reflects on unresolved tensions in AI governance and offers recommendations for forging a path toward a Just AI future.
As the glittering halls of the Palexpo buzz with the promises of agentic AI and innovations reshaping our physical world, a crucial question hangs in the air, one that resonates deeply from the vibrant, yet vulnerable, digital landscape of Africa: How do we approach AI trust and safety for just AI futures?
The optimism surrounding AI’s transformative potential is undeniable, yet beneath the surface lies a growing unease. We’re witnessing a proliferation of AI innovations at the company level with no clear accountability beyond voluntary trust and safety pledges. This is a precarious foundation.
As experts at the Summit cautioned: ‘You can’t truly test for bugs you haven’t seen before; our current safety verification methods mitigate for identifiable risk.’ This leaves us in a disquieting state where complex AI systems are seemingly waiting for a moment when the risk presents, rather than guaranteeing safety from inception.
For safe and Just AI, we must test for specific and concrete attributes of safety and trustworthiness. Without standardised attributes, safety verification at the AI innovation level remains inconsistent and wholly unreliable. If the library of attributes is not developed in adaptive ways, systems can pass stringent technical tests in controlled environments, only to fail spectacularly in the real world when confronted with unpredictable user interaction.
Preparedness reviews and responsible scaling policies from leading AI developers like OpenAI are a start, but they remain voluntary, reflecting internal characterisations of threats rather than standardised industry-wide protocols.
The challenge is magnified at the geopolitical level, where diverse priorities mean that concepts of AI risk or AI issues differ across continents. Myriad governance instruments amplify conceptual tensions and frustrate the implementation of governance processes to decisively address AI safety.
These unresolved tensions in AI governance pose fundamental questions for AI futures. The prospect, in particular, of superintelligence raises existential questions about maintaining human control and understanding how AI might evade it.
The path forward demands decisive action for outcomes-focused AI governance. Below are emerging recommendations:
- Principles-based approaches to AI governance must urgently meet governance in practice at the regulatory layer. While standards offer clarity for implementation, they cannot replace robust regulation.
- Transparency regulation is essential. For true collaborative risk appraisal, innovators must be open about faults as a precondition of relevant responses.
- For better AI outcomes, regulation can progress voluntary company safety policies to standardised, enforceable industry-wide safeguards overseen by independent National Safety Institutes working consultatively with local stakeholders. Generative AI models, for instance, can be subjected to mandatory rigorous safety and security tests at the national level before being made available to the public.
- Regulating content provenance through watermarking, explicit labelling, and embedded metadata is essential for humans to distinguish AI-generated content
- Adaptive regulation informed by research can contextualise AI innovation and capability with user experience, including safety experience.
- Integrate human rights and human values as AI governance red lines: While perfectly responsive technical risk mitigation is impossible, clearly articulating and enforcing human rights and human values as inviolable ‘red lines’ for AI innovation is achievable and imperative. This should go beyond human rights to inviolable human values. Beyond a right to privacy, how do we code for the independence of the user from AI decision-making?
- Prioritise Research & Development for contextualised risk clarification: Foster greater investment and impetus in R&D specifically focused on clarifying and contextualising AI risks, particularly those manifesting in the Global South and impacting marginalised groups both now and in the future. Empirical, scientific research provides a common language for identifying what works and what doesn’t.
- Streamline and harmonise AI governance initiatives: Actively work towards consolidating and making sense of the myriad of existing AI governance principles and emerging regulatory frameworks to reduce complexity, facilitate navigation, and improve compliance for all stakeholders.
- Mandate transparency in risk management: Require AI companies to publicly publish their risk management frameworks and provide systematic updates on the implementation of their risk mitigation measures. This allows the broader AI development community to share risk appraisals and enhance collective response capabilities.
- Establish international red lines and early warning indicators: Initiate a global effort to define what constitutes unacceptable AI-related risks in practice and develop practical early warning indicators. This must explicitly include risks present in the Global South and for marginalised groups.
- Scale global registration and identification of generative AI models: Accelerate ongoing efforts for the global registration and identification of generative AI models and agents to ensure traceability and accountability for their deployment.
- Enforce user protection by design at the application layer: Mandate that application providers build in default user protections, particularly regarding privacy. This means moving beyond relying on users to configure their settings, to application providers constructively protecting users.
- Institutionalise dynamic safety processes: Consider how the Seoul Commitments and similar frameworks that define intolerable risks and plans for managing them can be institutionalised for dynamic safety processes encompassing continuous testing, independent auditing, robust verification, ongoing monitoring, and transparent reporting throughout the entire AI lifecycle.
- Democratise AI governance: Leverage open foundations and shared resources for global governance knowledge. This is about allowing users sovereign ability to own and improve governance processes, and building literacy and incentives for this user-led governance layer.
