Preparing to Monitor the Impacts of Agents: Closing the Global Assurance Divide for Safe and Trusted AI

Abstract

The session wrapped up the India AI Impact Summit by linking the newly adopted Delhi Declaration to concrete actions for AI assurance, especially for autonomous “agentic” systems. After a ministerial keynote that outlined Singapore’s proactive sandbox and governance framework, a diverse panel examined what AI assurance means in practice, how standards, testing, and third‑party verification can be built, and why the split between well‑resourced “north” and under‑served Global South must be narrowed. Participants highlighted language diversity, infrastructure gaps, data‑supply‑chain dignity, and the need for interoperable, continuously‑monitored safety regimes. The rapid‑fire round distilled actionable commitments for multilateral bodies, global‑south interoperability, and Frontier‑model accessibility. The session closed with a call for shared, inclusive infrastructure to embed assurance throughout the lifecycle of agentic AI.

Detailed Summary

• Rebecca Finlay opened by positioning the panel within the Delhi Declaration adopted the previous day, stressing the need to translate high‑level commitments into scientific evidence and policy frameworks.
• Two new PAI papers were introduced:
  1. “Strengthening the AI Assurance Ecosystem” – a guide for national policymakers to embed assurance in industrial AI strategies.
  2. “Closing the Global Assurance Divide” – an analysis of how assurance challenges differ across the Global South and the Global North.
• QR codes and downloadable PDFs were promised; attendees were invited to approach any of the listed speakers for deeper discussion.

2. Minister Josephine Teo – Proactive Governance of Agentic AI

• Contextual shift: From reactive regulation to proactive preparation.
• Government as early adopter: Singapore runs an agentic‑AI sandbox with Google, allowing the public sector to “eat its own dog food” and learn safety implications before wider rollout.
• Model Governance Framework: A living document issued by the Singapore government to give enterprises concrete steps for responsible deployment; feedback is solicited continuously.
• Risk articulation: Autonomy yields productivity but also new failure modes (e.g., erroneous health or benefits advice).
• Assurance as competitive advantage: Companies that can demonstrate high safety assurance will differentiate themselves in the market.
• Three‑pillar assurance ecosystem:
  1. Testing – robust technical assessments that capture not only outputs but intermediate reasoning steps.
  2. Standards – agreed‑upon safety and reliability thresholds.
  3. Third‑party assurance providers – independent auditors to certify claims and expose blind spots.
• Call for collaboration: Singapore welcomes international partners to co‑design these components.

3. Panel Discussion – Perspectives on Global AI Assurance

3.1. ITU (Doreen Bogdan‑Martin) – Trust, Standards & Connectivity Gaps

• Trust as a bottleneck: While AI‑for‑Good use cases (healthcare, education, disaster response) proliferate, reproducibility across regions remains limited.
• From principle to practice: Standards must embed human‑rights, inclusivity, and security; the challenge is translating high‑level intent into enforceable technical specifications.
• Connectivity & language: About 2.6 billion people remain offline; lack of locally relevant content and services hinders adoption. AI can reduce friction (e.g., providing content in low‑resource languages), but must be paired with digital‑literacy programmes.
• Analogy to “leapfrogging”: The M‑Pesa mobile‑payment model shows potential for rapid adoption, yet no guarantee that AI will be used responsibly without proper governance and capacity building.

3.2. Google DeepMind (Owen Larter) – Agentic Systems, Protocols & Security

• What agents are: Autonomous systems that can plan, act, and orchestrate across multiple services (e.g., ordering dry‑cleaning without step‑by‑step user instructions).
• Emerging standards:
  • Agents‑to‑Agents Protocol – a universal language for agents to advertise capabilities and negotiate tasks.
  • Universal Commerce Protocol – enables secure transactions between agents and websites.
• Assurance standards: Need technical testing frameworks that quantify risk and verify behavior under diverse contexts.
• Security focus: Partnered with VirusTotal to scan agent‑downloaded “skills” for malware; highlighted credential‑sprawl risks when agents access email, banking, or other personal accounts.
• Computational accessibility: Introduced “Flash” models that are cheap and fast, arguing that low‑cost inference is essential for broad testing and real‑time monitoring.

3.3. African Language NLP (Vukosi Marivate) – Local Context & Data‑Supply Chains

• Masakane Research Foundation: Distributed network of thousands of African‑language researchers building NLP models for African languages.
• Non‑universal standards: Assurance frameworks designed in the Global North often ignore local linguistic diversity, leading to failed user experiences that can erode dignity.
• Data‑supply‑chain ethics: Emphasised the need for fair compensation, attribution, and consent for language data contributors; warned against “extractive” practices that ignore contributors’ rights.
• Capacity constraints: Policymakers in many Global South nations lack technical expertise and resources for system monitoring and compliance, making top‑down mandates insufficient.

3.4. Partnership on AI (Stephanie Ifayemi) – Mapping the Assurance Divide

• Six challenge domains (from PAI paper):
  1. Language diversity – evaluating models across 120 + Indian languages, 1 500 + African languages, and countless dialects.
  2. Risk profile variance – differing priorities (e.g., environmental assurance in Pacific islands vs. privacy in the US).
  3. Infrastructure – massive compute needs (e.g., 12 billion tokens → 19 500 GPU‑hours) create barriers for low‑resource settings.
  4. Documentation & disclosure – need for transparent model cards that reflect local contexts.
  5. Third‑party assurance capacity – scarcity of independent auditors in many regions.
  6. North‑South collaboration – ensuring Global South voices shape emerging standards (e.g., NIST/KC agent‑identity work).
• Strategic focus for the next year:
  • Incentive structures (e.g., insurance mechanisms) to motivate companies to invest in assurance.
  • Professionalisation – accreditation pathways for assurance practitioners and organisations.
  • Tiered assurance – aligning depth of testing with risk/impact of the use‑case (finance vs. medical).

4. Rapid‑Fire Q&A

Question	Speaker (Responder)	Key Points
Role of multilateral bodies (ITU) in inclusive assurance?	Doreen Bogdan‑Martin	Emphasised AI‑for‑Good as an inclusive “Davos‑of‑AI”; stresses practical pilots that feed into standards and policy; multilateral institutions must bridge theory and implementation.
How to make “test‑once‑comply‑globally” interoperable rather than exclusionary?	Vukosi Marivate	Argues contextual evaluations cannot be one‑size‑fits‑all; need repeatable, locally‑adaptable guidelines and data provenance checks; stresses the importance of user‑centric testing.
One concrete commitment Frontier Labs (FMF) should make?	Owen Larter	Commit to open, low‑cost model access plus multilingual benchmark suites; accelerate regional partnerships (e.g., with IIT Bombay) to improve performance on under‑served languages.
What should the global AI assurance landscape achieve in 12 months?	Stephanie Ifayemi	Four outcomes: (1) Robust incentive mechanisms (e.g., insurance); (2) Professional accreditation for auditors; (3) Standardised, multilingual evaluation suites; (4) Concrete North‑South standard‑making participation (e.g., comment on NIST agent‑identity paper).

5. Closing Reflections & Call‑to‑Action

• Stephanie Ifayemi synthesized the panel’s insights: Assurance must become an operational discipline, interoperable, and shared across sectors and geographies.
• Madhu Srikumar highlighted a tiered assurance model based on risk, reversibility, and autonomy levels of agents.
• Natasha Crampton (briefly introduced) signalled a forthcoming Microsoft perspective on responsible AI and the next steps for industry‑led assurance.
• The session concluded with an emphatic invitation to download the two PAI reports, join collaborative initiatives, and treat assurance as critical infrastructure that must be built collectively.

Key Takeaways

• Agentic AI raises new safety questions: Autonomy amplifies risk, demanding continuous, real‑time monitoring beyond pre‑deployment testing.
• Three‑pillar assurance ecosystem (testing, standards, third‑party verification) is essential; all three must evolve to handle dynamic, multi‑step agent behaviour.
• Global language diversity is a core technical hurdle: Benchmarks and evaluation protocols must support hundreds of languages and dialects to avoid reinforcing the north‑south divide.
• Infrastructure inequities matter: Compute‑heavy evaluation pipelines (e.g., billions of tokens) are prohibitive for many countries; low‑cost models and shared compute resources are needed.
• Standardisation must be inclusive: Institutions like ITU, NIST/KC, and regional AI safety institutes should co‑author standards and ensure Global South participation.
• Trust is both a market differentiator and a public‑good: Companies that can prove high assurance will gain competitive advantage, while governments need credible, independent certification to protect citizens.
• Rapid‑fire insights distilled actionable commitments: (i) Multilateral bodies to turn “principles → practice”; (ii) Interoperable “test‑once‑comply‑everywhere” frameworks must be locally adaptable; (iii) Frontier labs to improve accessibility of models and multilingual benchmarks.
• Professionalisation & incentives: Establish accreditation pathways for assurance professionals and insurance‑style incentives to drive industry uptake.
• Collaboration is non‑negotiable: No single entity can close the assurance divide; success hinges on shared infrastructure, open data, and joint standards‑writing across academia, industry, civil society, and governments.

---

Prepared for the AI Impact Summit (Delhi, 2024). All statements have been paraphrased for clarity while preserving the original meaning.

See Also:

• shaping-secure-ethical-and-accountable-ai-systems-for-a-shared-future
• navigating-the-ai-regulatory-landscape-a-cross-compliance-framework-for-safety-and-governance
• best-practices-from-the-international-network-for-advanced-ai-measurement-evaluation-and-science
• responsible-ai-for-bharat-building-trust-safety-and-global-leadership
• accelerating-indias-ai-growth-a-blueprint-for-indias-ai-success
• ai-agents-for-a-better-tomorrow-government-services-climate-action-and-resilient-infrastructure
• ai-innovators-exchange-accelerating-innovation-through-startup-and-industry-synergy
• building-sovereign-deep-tech-for-a-resilient-future-solutions-from-finland-and-india
• inclusion-for-social-empowerment
• responsible-ai-at-scale-governance-integrity-and-cyber-readiness-for-a-changing-world