Financial institutions face a critical new pressure: regulators now want them to open the black box and show exactly how their AI systems make decisions. The Financial Stability Board (FSB), an international body coordinating financial regulation across major economies, published a consultation report on June 10, 2026, proposing 12 sound practices for how banks should responsibly adopt artificial intelligence. The stakes are high. These practices are not yet binding, but they will likely become benchmarks for supervisory examinations once finalized.

The challenge for compliance teams is thornier than it first appears. The FSB's new guidance overlaps significantly with prior frameworks, but introduces subtle inconsistencies in terminology and definitions that create ambiguity about what regulators actually expect. For financial institutions, the window to shape these standards is closing fast: comments are due by July 22, 2026.

What Are the FSB's 12 Sound Practices?

The FSB grouped its recommendations into two broad clusters. The first four practices focus on organization-wide AI governance, including risk appetite frameworks, governance and accountability structures, and how banks integrate AI risks into their broader risk management frameworks. The remaining eight practices address the full lifecycle of AI systems, from selection and data governance through to explainability, performance monitoring, and human oversight.

Among these, Sound Practice 8 on explainability and transparency stands out as particularly contentious. The FSB's definition diverges from how the Organization for Economic Cooperation and Development (OECD) and the European Union define the same concept. This creates interpretive uncertainty for banks operating across multiple jurisdictions. Similarly, practices around materiality assessment, performance management, and human oversight introduce conceptual restatements with multiple competing definitions, elevating ambiguity risk.

Why Is Interpretability Suddenly a Regulatory Priority?

The FSB identified the same core AI risks back in 2017: third-party dependencies, market correlations, lack of interpretability and auditability, cyber risk, governance gaps, and bias. But the language has evolved. What the FSB called "observations" in 2017 became "vulnerabilities" in 2024, and now appears as "sound practices" in 2026. The underlying risks haven't changed materially, but regulators are tightening the screws on how banks must demonstrate they understand and control these risks.

Interpretability matters because it enables auditing. When a bank deploys an AI system to approve loans, detect fraud, or manage trading risk, regulators want to know not just whether it works, but why it works. A model that achieves 95% accuracy on historical data but operates as a black box is far riskier than a slightly less accurate model whose decision logic can be inspected and validated. This shift reflects a broader recognition that AI systems in finance can amplify systemic risk if they fail in correlated ways across institutions.

How Are Researchers Advancing AI Interpretability?

Meanwhile, the academic and commercial AI community is racing to solve the interpretability problem from a different angle. Recent research by Posner, Lei, and Schölkopf formalized the concept of "Mechanistic World Models," which promise to make AI systems more transparent by placing reusable mechanisms at the heart of computation. Rather than treating AI as a black box that makes predictions, mechanistic interpretability enables researchers to inspect each variable's causal role and understand how the system reasons.

This approach is gaining traction in industry. Google DeepMind's Genie model, which sports approximately 11 billion parameters, powers high-fidelity interactive video rollouts and is being marketed as a Scientific World Model. Wayve's GAIA family reaches nine billion parameters for autonomous driving simulation. Both companies mention planned tooling for interpretability audits in internal reports, signaling that mechanistic interpretability is becoming a competitive differentiator.

What Challenges Still Block Progress?

Despite the momentum, significant obstacles remain. Jointly learning variables, mechanisms, and structure still lacks a scalable algorithm. Partial observability distorts signal, complicating causal reasoning and explainability efforts. Non-Markovian dynamics, which violate assumptions baked into many video prediction datasets, further complicate the picture. Current world models often overfit to video artifacts, limiting generalization beyond the narrow domains where they were trained.

No end-to-end open implementation of a full Mechanistic World Model exists today. This means that while the theory is advancing rapidly, the engineering to translate it into production systems remains incomplete. Researchers therefore explore active experiment selection to resolve ambiguity through intervention, but this approach has not yet scaled to real-world financial systems.

How Can Banks Prepare for the New Standards?

• Audit Existing Models Now: Before the FSB standards become binding, financial institutions should conduct targeted reviews of their current AI systems to identify which ones lack explainability. This includes models used for credit decisions, fraud detection, and market risk assessment.
• Respond to the FSB Consultation: The comment period closes July 22, 2026. Banks should submit targeted feedback highlighting where the FSB's new definitions conflict with existing local, sectoral, or firmwide requirements, and propose clarifications that reduce ambiguity.
• Invest in Interpretability Tooling and Skills: Companies deploying decision-making agents now face sharper regulator scrutiny regarding model transparency. Firms should train teams in mechanistic interpretability and causal reasoning practices, and evaluate whether internal stacks qualify as Scientific World Models before green-lighting deployment.
• Align with Academic AI for Science Groups: Partnerships with academic researchers working on mechanistic interpretability can shorten technology transfer cycles and help banks understand how to implement transparent mechanisms in production systems.

The FSB's core AI risk framework dates back nearly a decade, and some building blocks trace to 2013. Yet the regulatory language has drifted, creating redundancy in some areas and unhelpful ambiguity in others. For financial institutions, the time for action is now. The open comment period is a critical window for market participants to shape the FSB's approach, rather than merely acquiescing to recommendations that may fail to improve risk outcomes or that may duplicate or conflict with existing requirements.

The broader message is clear: regulators are no longer willing to accept AI systems they cannot understand. Banks that invest early in mechanistic interpretability and transparent decision-making will find themselves better positioned for compliance, while those that delay risk costly setbacks. As one industry observer noted, boards are now asking whether internal stacks qualify as Scientific World Models before green-lighting deployment. Transparent mechanisms safeguard trust, compliance, and adaptability, making investment in skills and tooling non-negotiable.