Closing the AI Trust Gap: The Case for Independent Certification for Trustworthy AI

This paper addresses the "trust gap" in the artificial intelligence industry—a situation where companies invest in internal safety and ethics, yet fail to provide the public, regulators, or investors with a reliable way to verify that their systems are truly trustworthy. The authors argue that while "Responsible AI" focuses on internal processes, it has not created a market that rewards companies for building safe, beneficial systems. To bridge this gap, the paper proposes a new system of independent, outcome-oriented certification that would allow organizations to prove their systems deliver real-world value, making trustworthiness a measurable and commercially rewarded standard.

The Problem: Why Current Efforts Fall Short

The authors identify three structural failures that prevent the current AI landscape from building public trust. First, the market cannot distinguish between companies that are genuinely trustworthy and those that are merely performing "responsible" marketing, leading to a "market for lemons" where high-quality efforts are not rewarded. Second, current evaluation methods focus on isolated model outputs—like benchmark scores—rather than the actual, long-term impact of AI systems when used in real-world, sociotechnical settings. Third, the current governance ecosystem is almost entirely focused on avoiding harm rather than demonstrating that an AI system provides a positive benefit to society.

The Distinction Between Responsible and Trustworthy AI

The paper draws a critical line between "Responsible AI" and "Trustworthy AI." Responsible AI is described as a process-based approach—the internal steps a company takes to manage risk. However, processes alone do not guarantee outcomes. Trustworthy AI, by contrast, is a relational property earned through evidence of real-world performance. The authors use the analogy of the automotive industry: we trust a car not just because of the manufacturer's internal processes, but because the vehicle has passed independent safety tests and certifications. The authors argue that AI needs a similar, external layer of infrastructure to prove that systems behave as promised over time.

A Proposal for Independent Certification

To fix these issues, the authors propose an independent, outcome-oriented certification framework. Unlike current voluntary disclosures, which can be costly and risky for companies to provide, this certification would serve as a standardized signal. By integrating a governance baseline with verified evidence of positive outcomes, this framework would allow stakeholders to compare AI systems based on their actual impact. The goal is to move beyond simple compliance and create a market mechanism where safety and benefit are recognized as competitive advantages, ultimately helping to align corporate incentives with the public interest.

Limitations of the Current Landscape

The authors note that while existing frameworks like the NIST AI Risk Management Framework and the EU AI Act provide important foundations, they are currently insufficient on their own. The primary limitation is the lack of institutional infrastructure to turn audit findings into meaningful, public-facing accountability. Furthermore, as AI models become more capable, they are increasingly able to "game" pre-deployment tests, meaning that one-off checks are no longer enough. The authors emphasize that any successful certification regime must be continuous, monitoring systems throughout their lifecycle to ensure they remain trustworthy as they are deployed and used in the real world.