An UNESCO report titled “Pathways on Capacity Building for AI Supervisory Authorities” highlights the urgent need for enhanced oversight in the rapidly evolving landscape of artificial intelligence (AI). Produced in partnership with the Dutch Authority for Digital Infrastructure (RDI) and backed by the European Union’s Technical Support Instrument, the report synthesizes discussions from the first UNESCO Expert Roundtable on AI Supervision held in Paris in May 2025. It incorporates insights from various organizations, including the Tony Blair Institute, EUSAiR, and Brazil’s Data Protection Authority (ANPD).
The document underscores a shared concern: AI technologies are proliferating across multiple sectors such as credit, healthcare, hiring, and education, yet the institutions tasked with their regulation are lagging. Unlike traditional technologies, AI systems learn from data and adapt over time, making conventional regulatory methods—characterized by fixed rules and one-time approvals—largely ineffective. As a result, supervisory authorities face dual challenges: a technical one involving the dynamic nature of AI systems and an institutional one regarding oversight capabilities.
The report explains that many regulators initially turned to oversight models from sectors like aviation and pharmaceuticals, which are based on stable standards and predictable risks. However, AI’s behavior often defies such predictability, as the intricacies of its operations cannot be fully understood through code inspection alone. The report further critiques the optimism around “explainable AI,” suggesting that attempts to simplify complex models have not yielded significant benefits for regulators. Consequently, the focus is shifting to establishing institutions capable of interpreting AI behaviors in context, rather than fully decoding them.
From Compliance to Continuous Oversight
The concept of “interpretative supervision” forms the crux of the report’s recommendations. This approach shifts oversight from static compliance checks to an ongoing process of observation, learning, and judgment. Supervisors are encouraged to ask pertinent questions: Is an AI system generating biased outcomes? Is its performance evolving over time? Are new risks emerging that were unknown at the time of its deployment?
To operationalize this new approach, the report introduces the OBSERVE framework, primarily developed by the Tony Blair Institute. This framework advocates for the establishment of dedicated observatory units among regulators, employing real-time monitoring tools, and leveraging external expertise. By focusing on the collection of evidence from past incidents, authorities can transition from reactive responses to proactive identification of potential issues.
Another key aspect discussed in the report is the use of AI regulatory sandboxes, environments that permit the testing of AI systems under regulatory oversight. The report clarifies that these sandboxes are not merely loopholes for companies but serve as valuable learning tools for regulators. By observing AI systems in practice, regulatory bodies can better understand the associated risks, clarify legal applications, and refine future regulatory frameworks.
Within the European Union, the report details how the EU AI Act mandates Member States to establish such sandboxes and highlights projects like EUSAiR aimed at coordinating sandbox initiatives across nations. Successful implementation of these sandboxes requires integration with broader innovation and testing infrastructures and must function as ongoing processes, not isolated experiments.
Brazil’s experience with regulatory sandboxes provides critical insights into the potential benefits and challenges. The Brazilian Data Protection Authority has utilized sandboxes not just for testing AI systems but for reforming how public institutions operate. The report notes the bureaucratic resistance often encountered, driven by fears of legal repercussions and institutional risk. Overcoming this resistance necessitates training, legal clarity, and a cultural shift within institutions.
The final takeaway from the report emphasizes that effective AI governance hinges on the establishment of robust institutions. Legal frameworks and ethical guidelines, including UNESCO’s own Recommendation on the Ethics of AI, will be futile unless supervisory authorities possess the requisite skills, tools, and confidence to enforce them. Strong supervision demands not only technical knowledge but also cross-sector collaboration, a willingness to learn, and adaptability in the face of technological change.
By drawing on real-world experiences from Europe and Latin America, the report illustrates that effective supervision can coexist with innovation. When executed correctly, oversight can mitigate uncertainties, safeguard public interests, and foster responsible technological advancement. Ultimately, capacity building is no longer a secondary concern but the central challenge for governments attempting to regulate artificial intelligence in a manner that serves the public good.
See also
Congo Launches Seminar with Global Experts to Shape Digital Regulation and AI Governance
Omer Tene Discusses Key Digital Policy Trends in Privacy and AI for 2026
Florida’s Sam Garrison Forecasts Heated AI Regulation Debate Amid Property Tax Cuts
Fairfax County School Board Approves $2.5M Renovation and Expels Two Students for Weapons
US AI Regulations Emerge: $500B Investment Fuels New Ethical Frameworks Across Sectors
