DNV Outlines Foundations for Achieving Trustworthy AI

Key Takeaways

• AI risk requires continuous, adaptive assurance throughout lifecycle
• Modular risk models break complex AI systems into manageable parts
• System model must capture AI, humans, digital, physical interactions
• Evidence‑linked claims provide auditable proof of safety
• Board‑level risk pressure drives adoption of AI assurance standards

Summary

DNV released a position paper outlining how traditional risk‑management principles can be adapted to assure AI‑enabled systems in safety‑critical industries. The research emphasizes a continuous, lifecycle‑wide assurance model that captures the full AI ecosystem, from data and algorithms to human and physical interactions. Core foundations include a comprehensive system model, modular risk decomposition, evidence‑linked safety claims, and adaptive monitoring as AI evolves. DNV is already partnering with firms to apply these methods, positioning the framework as a practical pathway for trustworthy AI adoption.

Pulse Analysis

DNV’s latest research bridges a gap that has long challenged AI developers in high‑risk environments: translating decades‑old safety assurance practices to the fluid, data‑driven nature of modern machine learning. By leveraging its heritage in maritime and energy risk management, DNV demonstrates that the same probabilistic safety analyses used for offshore platforms can be re‑engineered to evaluate emergent AI behaviours. This continuity not only preserves institutional knowledge but also offers a familiar language for regulators and insurers who are still grappling with AI‑specific liabilities.

The paper’s four pillars—comprehensive system modelling, modular risk decomposition, evidence‑linked safety claims, and continuous context‑aware assurance—form a pragmatic toolkit for organizations seeking to embed trustworthiness into AI lifecycles. A holistic system model forces engineers to map interactions between algorithms, operators, and physical assets, exposing hidden failure modes that isolated component testing would miss. Modular risk assessment then slices these complexities into bite‑size units, enabling targeted mitigation and clearer accountability. Linking safety claims to verifiable evidence creates an audit trail that satisfies both internal governance and external certification bodies, while real‑time monitoring ensures that model updates or data drift do not erode previously established confidence.

For corporate boards, the implications are immediate. As insurers tighten underwriting criteria around AI‑driven operations, demonstrable assurance becomes a differentiator that can lower premiums and protect against litigation. Moreover, emerging regulations in the EU and United States are beginning to codify AI risk management standards, making DNV’s framework a potential baseline for compliance. Early adopters stand to gain competitive advantage by reducing time‑to‑market for AI solutions while safeguarding critical services. In the longer term, widespread adoption of these assurance practices could set industry‑wide expectations for transparent, resilient AI, fostering a market where trust is engineered rather than assumed.