Executive Summary:
- Trustworthy AI governance has moved beyond principles to practice: Hiroshima showed that trust is no longer treated as a normative input, but as an outcome of governance infrastructure: continuous evaluation, incident handling, reporting, and cooperation across institutions and jurisdictions.
- AI Safety Institutes (AISIs) are emerging as critical coordination nodes: AISIs are not regulators or research labs, but trusted intermediaries that translate technical evaluation into governance-relevant signals, enable shared learning, and support interoperability
- Agentic AI is the stress test forcing governance to become continuous and system-level: As AI systems plan, act, and adapt over time, risk shifts from outputs to workflows and runtime behavior. This makes static compliance insufficient and elevates evaluation, monitoring, and remediation as core governance functions.
- Global AI governance is converging on cooperation without uniformity: Rather than harmonized rules, the emerging model emphasizes complementarity: shared artifacts (evaluations, reporting, taxonomies) mapped across EU, US, and APAC approaches, with growing attention to capacity, inclusion, and legitimacy in the Global South ahead of the India AI Impact Summit.
The Hiroshima Global Forum for Trustworthy AI, co-hosted by Japan’s AI Safety Institute and the Cabinet Office, was convened, between January 15-16, as an international, multi-stakeholder forum to examine emerging approaches to AI safety, security, and trustworthiness, with a particular focus on how national AI safety institutions, governments, industry, and research actors can coordinate governance practices across jurisdictions.
The Forum did not feel like a summit designed to announce a breakthrough. It felt more like a working session for a system already under construction; one that many participants recognized as necessary, imperfect, and still fragile.
Rather than converging around a single vision of trustworthy AI, the Forum revealed something more instructive; namely, that trust is no longer being debated as a principle, but assembled through practice. Across government, industry, and research institutions, the discussion repeatedly returned to concrete questions of coordination: namely, who evaluates what, how results are shared, how failures are handled, and how different governance approaches can coexist without collapsing into fragmentation?
This shift matters because it marks a transition in global AI governance. The era of agreeing on values (fairness, transparency, accountability) has not ended, but it is no longer the frontier. The frontier is now how those values are made operational across jurisdictions, sectors, and stages of the AI lifecycle, especially as systems become more autonomous, more interconnected, and more difficult to control through static rules. Hiroshima offered a snapshot of that transition in motion.
Trust as an Outcome, Not an Input
One of the most consistent undercurrents at the Forum was a reframing of trust itself. Trust was rarely discussed as something that could be demanded or declared. Instead, it emerged as something earned indirectly, through reliable behavior, predictable responses to failure, and visible accountability mechanisms.
This was especially apparent in discussions that linked trust to adoption and investment. Trust was not framed as a social license layered on top of innovation, but as the condition that allows innovation to scale. When trust fails, deployment stalls, procurement slows, and political pressure intensifies. What stood out was that trust was repeatedly tied to processes, not promises: evaluation pipelines, reporting structures, incident handling, and cooperation mechanisms. Trust, in this framing, becomes the by-product of governance infrastructure functioning well enough that actors (regulators, policy-makers, industry and end-users) can rely on it even when systems behave unpredictably and in a trustworthy fashion.
Why the AISI Network Became a Focal Point
Against this backdrop, one of the most consequential elements of the Forum was the growing prominence of AI Safety Institutes (AISIs); not as regulators, and not as research labs in the traditional sense, but as coordination nodes in an increasingly complex governance ecosystem.
What became clear in Hiroshima is that AISIs are beginning to occupy a distinctive role that neither governments nor industry can easily fill alone:
- They sit close enough to technical evaluation to engage with real system behavior.
- They sit close enough to policy to translate findings into governance expectations.
- And they are increasingly networked internationally, allowing methods, taxonomies, and lessons to travel without requiring legal harmonization.
The value of AISIs was not framed in terms of authority, but in terms of function. Multiple speakers implicitly converged on the idea that frontier risks will often be detected first by industry (because industry is behind or operates the systems in question) but that risk maturation, boundary-setting, and shared understanding require institutional spaces that are trusted across sectors.
The AISIs network emerged as a clear point of reference as a forum where evaluation, benchmarking, incident analysis, and research could be aggregated, contextualized, and communicated without collapsing into either regulatory capture or purely academic abstraction.
Importantly, AISIs were not presented as monolithic. The diversity of approaches (i.e. Japan’s evaluation-heavy model, Singapore’s tooling and sandbox emphasis, Korea’s institutional backbone, Canada’s methodological rigor) was not treated as a problem to be solved, but as a feature to be coordinated.
Agentic AI as a Governance Stress Test
Another element that emerged in Hiroshima was how decisively discussions have shifted toward agentic AI. This was not treated as a distant frontier concern, but as a near-term governance stress test. Agentic systems (which decompose goals, plan, use tools, act, and adapt over time) undermine many of the assumptions that earlier AI governance relied on. Risk is no longer confined to outputs. It unfolds over sequences of actions, across tools, and within socio-technical systems that include humans in the loop.
Participants repeatedly highlighted failure modes that do not fit cleanly into existing categories: cascading hallucinations, loss of control in long-running workflows, unintended tool use, and emergent behavior that only appears at runtime. These risks cannot be fully mitigated through pre-deployment audits or static documentation. What emerged instead was a consensus, often implicit, that governance must become continuous, system-level, and as close to the use-case. Evaluation must happen not just before deployment, but during operation. Incident handling must be normalized, not treated as exceptional. And responsibility must be distributed across organizations, not assigned solely to models.
This is another reason AISIs matter.
As agentic systems blur boundaries between development and deployment, between safety and security, and between technical and organizational risk, governance increasingly depends on shared evaluation practices and learning loops rather than rigid rules. Highlighting this, on January 22nd, Singapore published its Model AI Governance Framework for Agentic AI which gives organisations a structured overview of the risks of agentic AI and emerging best practices in managing these risks (by identifying, controlling, and deploying agentic AI responsibly through risk-bounded use cases, human accountability, lifecycle controls, and end-user transparency.).
Interoperability Reconsidered: From Uniformity to Cooperation
This diversity feeds directly into one of the most contested concepts at the Forum, namely interoperability. Interoperability was everywhere in the language of the summit, but rarely uncontested. What emerged instead was a more nuanced conversation about what kind of interoperability is desirable, and at which layer.
Rather than pushing toward uniform global rules (an increasingly implausible goal in the current policy landscape we are witnessing) many participants leaned toward a model of complementarity and compatibility. In this model, different jurisdictions retain distinct governance approaches, but align around shared artifacts: comparable evaluations, interoperable reporting structures, mutually intelligible risk categories, and transparent processes.
This reframing matters because it acknowledges political reality without abandoning coordination. It also reflects a deeper recognition: governance friction is not always a failure. From a human rights and Global South perspective, friction can protect against asymmetric interests being silently embedded in universal standards.
What interoperability increasingly seems to mean, in practice, is the ability to map and translate. Additionally, this could be mirrored in framing how to understand how different regimes address similar risks, how evidence can be reused across contexts, and how accountability can travel even when laws do not. This is where the AISI network becomes strategically important. AISIs offer a venue for this translation work to happen at the technical and methodological level, rather than being forced prematurely into legal alignment.
Transparency, Reporting, and the Limits of Voluntarism
The Hiroshima AI Process (HAIP) provided a concrete case study of how these ideas are being operationalized internationally. What stood out was not the novelty of HAIP’s principles, but the architecture of its reporting framework.
By translating broad commitments into structured reporting pillars (risk identification and mitigation, governance, transparency, incident management, provenance, investment, and societal interests) HAIP turns voluntary alignment into something observable. The emphasis is explicitly on monitoring and learning, not enforcement.
What emerged in discussions is that voluntarism is not inherently weak. It becomes weak only when it lacks structure, visibility, and follow-through. In contrast, structured transparency can exert real governance pressure through peer comparison, procurement expectations, and reputational signaling, without requiring binding law.
This model is particularly relevant in a world where not all countries can, or want to, adopt the same regulatory regimes. HAIP’s architecture suggests a way to build shared accountability without uniform legislation, a feature that becomes increasingly important when considering Global South participation.
Internal Cooperation Before Global Convergence
One of the quieter but more important insights from Hiroshima was that internal cooperation often precedes international cooperation.
Across regions, participants emphasized the difficulty of aligning governance even within governments, let alone across borders. Fragmentation between ministries, regulators, and technical agencies remains a core challenge. AISIs, again, were often described as a way to address this, by centralizing evaluation expertise and serving as internal coordination hubs.
This internal dimension is critical when thinking about the Global South and the upcoming India AI Impact Summit. Many countries face governance capacity constraints that make fragmented approaches untenable. For them, the question is not how to harmonize with every global framework, but how to build internal coherence first, and then connect outward selectively.
India’s role here is particularly significant. As a major AI developer and deployer with deep ties to both Western and APAC ecosystems, India is positioned to foreground questions of capacity, inclusion, and development that are often peripheral in transatlantic debates. The Summit’s emphasis on impact (rather than abstract safety) in domain-specific applications and sectors reflects this shift.
EU-US-APAC: Cooperation Without Convergence
Hiroshima also clarified the contours of emerging cooperation between the EU, the US, and APAC nations.
The EU continues to anchor the regulatory imagination, translating risk into law and shaping global compliance behavior through market power. The US remains the engine of frontier capability and deployment, shaping de facto standards through tooling, ecosystems, and scale. APAC nations (particularly Japan, Singapore, and Korea but also in ASEAN) are increasingly positioning themselves as operational governance innovators, building testing infrastructure, assurance sandboxes, and institutional capabilities that bridge policy and standards with practice.
What Hiroshima suggested is that cooperation among these regions will likely occur below the level of law. Shared evaluations, joint research agendas, interoperable reporting structures, and AISI-to-AISI collaboration may prove more effective (and more politically feasible) than attempts at formal regulatory and standards alignment.
This model of cooperation respects difference while enabling learning. It also creates space for the Global South to engage through participation in shared infrastructure, rather than being forced to adopt externally defined regimes wholesale.
Toward a More Grounded Vision of Trustworthy AI
Taken together, the Hiroshima Global Forum painted a picture of trustworthy AI governance that is less aspirational and more efficient. Trust is no longer treated as something we can design once and deploy everywhere. It is treated as something that must be maintained through cooperation, tested through evaluation, repaired through remediation, and legitimized through inclusion.
The added value of the AISI network lies precisely here; not in solving governance, but in making governance possible at scale, across systems that are evolving faster than any single institution can manage alone. As attention shifts toward New Delhi and the Global South, the test for trustworthy AI governance will not be whether it sounds coherent, but whether it can accommodate diversity (of risk, of capacity, of context) without collapsing into either fragmentation or dominance.
Again, the Hiroshima Forum did not offer final answers. But it stressed that the future of trustworthy AI will be built less through declarations, and more through the quiet, difficult work of coordination.
Evaluation results, risk categorizations, incident responses, reporting obligations, and organizational controls are multiplying across jurisdictions, frameworks, and institutions. What is missing is not intent, but connective infrastructure that allows these elements to work together without forcing artificial uniformity.
Platforms like Credo AI have a concrete role to play here: translating diverse governance expectations into operational workflows; mapping evidence across regimes such as the EU AI Act, HAIP, NIST, and ISO; and enabling organizations to demonstrate trustworthiness through continuous, auditable practice rather than static declarations. In a world moving toward agentic systems and pluralist governance, this kind of orchestration is not optional. To the contrary, it is the practical foundation of trust.
The priority is building coherent, capacity-appropriate governance foundations before engaging internationally. India is uniquely positioned to shape this next phase, grounding trustworthy AI in impact, inclusion, and real-world deployment. Stay tuned for more take-aways from our Credo AI presence and activities during the course of the upcoming India AI Impact Summit, where these questions will move from discussion to action.
DISCLAIMER. The information we provide here is for informational purposes only and is not intended in any way to represent legal advice or a legal opinion that you can rely on. It is your sole responsibility to consult an attorney to resolve any legal issues related to this information.
