Page 27 - The Annual AI Governance Report 2025 Steering the Future of AI

The Annual AI Governance Report 2025: Steering the Future of AI

to “trailing-edge” work—high-level management and process standards for well-established
problems—rather than trying to codify every frontier risk.
84
One practical way to address this “standardisation gap” would be to adopt ITU’s pre-
standardisation focus group model , which has been used in areas such as health and disaster
85
management in collaboration with other UN agencies. This could involve establishing a joint
ITU–ISO/IEC group to develop a reference architecture, risk taxonomy, and sandbox test
framework. Meanwhile, IEEE could develop additional ethics-by-design guidance.

Standards for AI Agents: Agentic AI will require new standards that go beyond today's model-
centric norms. No formal standard yet specifies agent-to-agent communication protocols,
secure tool APIs, or guardrails for autonomous, self-modifying behaviour. Early prototypes,
such as Microsoft's open Agent2Agent (A2A) protocol and the community-led Model
86
Context Protocol (MCP) , demonstrate how multi-agent workflows and context exchange
87
could function. However, they remain ad hoc specifications outside the remit of any formal
standards development organisation. Safety researchers warn that existing test methods do
not capture the emergent risks of agents. The newly published Multi-Agent Emergent Behavior
(MAEBE) framework is one of the first attempts to measure collective behaviours in the absence
of harmonised evaluation methods.
88

4.3 Ethical AI Frameworks

UNESCO Recommendation on the Ethics of Artificial Intelligence: On 23 November 2021,
UNESCO adopted the Recommendation on the Ethics of Artificial Intelligence – the first multilateral
agreement on AI ethics. It commits all 194 Member States to the principles of protecting
human rights, transparency, fairness, and human oversight. To turn these high-level values
89
into practice, UNESCO created a Readiness Assessment Methodology (RAM), and according
to the agency’s Social Sciences directorate, it has 'worked with nearly 60 countries — largely in
Africa, the Caribbean and Latin America — to run baseline diagnostics and draft action plans'.
90
While implementation efforts are still in early stages, and some concepts (such as “fairness” and
“accountability”) remain open to context-specific interpretation , the Recommendation serves
91
as a valuable ethical framework. As metrics and regulatory pathways continue to evolve, the
Recommendation provides important guidance and supports international convergence on AI
ethics, especially in contexts where formal regulatory regimes are still emerging.

OECD AI Principles: Adopted by OECD ministers on 22 May 2019 and revised on 3 May 2024 to
address the risks posed by generative AI, the OECD Recommendation on Artificial Intelligence

84 Huw R. and Ziosi M. (2025, June 9) Can we standardise the frontier of AI? Oxford Martin AI Governance
Initiative.
85 International Telecommunication Union. (2025, May 5). AI/ML (Pre-) Standardization - AI for Good. AI For
Good.
86 Arenas, Y., & Brekelmans, B. (2025, May 22). Empowering multi-agent apps with the open Agent2Agent
(A2A) protocol | The Microsoft Cloud Blog.
87 Model Context Protocol. (2025, March 26). Specification - model context protocol.
88 Erisken, S., Gothard, T., Leitgab, M., & Potham, R. (2025, June 3). MAEBE: Multi-Agent Emergent Behavior
Framework. arXiv.org.
89 https:// www .unesco .org/ en/ artificial -intelligence/ recommendation -ethics.
90 UNESCO. (2023). Readiness assessment methodology. A tool of the Recommendation on the Ethics of
Artificial Intelligence.
91 AllahRakha, N. (2024). UNESCO's AI Ethics Principles: Challenges and Opportunities. International Journal
of Law and Policy, 2(9), 24–36.

18