Independent Assurance for Safe and Reliable AI
How can people reasonably rely on consequential AI systems they cannot independently evaluate?
Artificial intelligence (AI) presents a problem familiar to professionals in the financial services industry: individuals, businesses, and governments increasingly rely on systems whose operation, results, and safety they cannot readily evaluate.
The rapid pace of AI development make it imperative that systems to be quickly be developed and deployed that allow developers, users and society at large to:
- Understand the extent to which they can rely on AI systems and
- Determine whether consequential AI systems are safe.
The question is not simply whether AI should be encouraged, restricted, or regulated. It is also:
What evidence should justify reliance on a particular AI system, and who should be trusted to evaluate it?
My proposal is to adapt established principles of independent assurance to artificial intelligence—creating a structure through which qualified, independent professionals can evaluate AI systems, outputs, deployments, and governance controls against defined criteria and provide users with a credible basis for reasonable reliance. This page provides a summary of a workable approach. Details are provided in an article and congressional summary available via these links:
Full article — PDF
Toward Safe and Reliable AI: How Independent Assurance Can Address AI Risks Without Sacrificing Its Benefits
Congressional summary — PDF
Independent Assurance for Safe and Reliable Artificial Intelligence
A Workable Approach to AI Assurance
- AI developers and deploying organizations are responsible for the AI systems they create, including appropriate internal testing, monitoring, documentation and controls.
- Public authorities or recognized standard setters should establish (or recognize) criteria for consequential AI applications.
- Qualified independent assurance providers would examine the relevant evidence, test the work performed internally, conduct additional procedures where necessary, and report conclusions regarding safety and reliability.
The objective is not to declare that an AI system is universally “safe,” “truthful,” or “unbiased.” An assurance conclusion would identify the particular system or application evaluated, its version and configuration, its intended use and operating conditions, the criteria applied, the level of assurance obtained, and important limitations. Material changes could require renewed evaluation.
What Would Be Assured?
The proposal distinguishes two fundamental assurance objectives.
Reliability
For a specified use and operating conditions, does the AI produce information or decisions that meet defined criteria for accuracy, evidentiary support, completeness, calibration and relevant forms of fairness?
Reliability assurance could examine such matters as whether citations genuinely support the propositions attributed to them, whether sources are sufficiently authoritative and relevant, whether important contrary evidence has been considered, whether facts are distinguished from inference or speculation, and whether the confidence expressed is appropriate to the underlying evidence.
Safety
Can the specified AI system and deployment operate within defined risk tolerances, considering its capabilities, autonomy, access, safeguards, monitoring and potential consequences?
Safety evaluation becomes particularly important as AI systems acquire greater capability, more autonomy and access to consequential digital or physical systems. Assurance requirements could become progressively more demanding as those factors increase.
A third subject is also important: AI governance and management controls. Organizations increasingly need credible ways to evaluate how AI is approved and used internally, how confidential information is protected, what access AI systems receive, how activity is logged and reviewed, and how incidents and material changes are handled.
A Layered Assurance Structure
No single institution needs to perform every function. A practical system could combine:
- externally established or recognized criteria;
- developer controls, testing and internal safety evaluation;
- governmental technical evaluation where appropriate;
- independent nongovernmental technical evaluation and research;
- qualified independent external assurance;
- oversight of the organizations and professionals providing assurance; and
- consequences when required criteria or assurance requirements are not satisfied.
These functions should be complementary rather than competing. Independent providers could evaluate and use reliable work performed by developers, reperform selected procedures, investigate exceptions and concentrate additional independent testing where risk or uncertainty is greatest.
Independent assurance produces credible evidence. Governmental authorities determine when that evidence is required and what consequences follow from it.
AI Will Be Part of the Assurance Process
The scale of AI evaluation may itself require artificial intelligence.
A purpose-built assurance AI could decompose outputs into assertions, examine evidentiary sources, independently search for contradictory evidence, compare system behavior across controlled variations, conduct large numbers of adversarial safety tests, identify anomalies and preserve detailed evidence of the procedures performed.
Such systems could perform evaluation at a scale, speed and frequency that would often be impractical for human teams alone.
But AI performing the procedures does not eliminate institutional accountability. People and organizations would still need to establish or recognize criteria, determine acceptable risk, evaluate whether the evidence is sufficient, oversee the assurance process and take responsibility for the conclusion that is reported. The assurance systems themselves would also require testing, monitoring and quality controls.
Assurance as an Enabling Mechanism
Independent assurance should not be viewed only as a way to constrain AI.
Credible assurance could also help trustworthy AI systems earn confidence. Businesses may be more willing to integrate AI, governments to permit consequential uses, insurers to accept risks and users to rely on systems when there is credible independent evidence that specified reliability and safety criteria have been satisfied.
Organizations that currently limit AI use because they cannot adequately evaluate confidentiality, security, reliability or governance risks may be able to expand its use when appropriate standards and assurance mechanisms provide a practical way to manage those risks.
Trust can itself become an economic asset.
The purpose of the proposal is therefore not to choose between innovation and safety. It is to develop an institutional structure capable of distinguishing AI systems and deployments that deserve greater confidence from those requiring additional safeguards or evidence.
Publications
Toward Safe and Reliable AI
How Independent Assurance Can Address AI Risks Without Sacrificing Its Benefits
Bruce Bleakman, CPA — September 2026
The full 23-page article develops the assurance concept, distinguishes reliability from safety, examines how AI itself could perform assurance procedures, considers lessons from financial reporting and gaming regulation, reviews emerging institutional approaches, and discusses qualification and oversight of assurance providers.
Full article — PDF
Congressional Summary
Independent Assurance for Safe and Reliable Artificial Intelligence
A shorter policy-oriented summary describing the proposed assurance architecture and how independent assurance could complement developer testing, governmental technical evaluation, standards development and regulatory authority.
Congressional summary — PDF
Updates
Update No. 1 — Transluce and the Emerging Independent AI-Evaluation Ecosystem
September 28, 2026
The original article distinguishes developer testing, governmental evaluation and independent external assurance. Developments since the article was circulated suggest that another complementary function deserves particular attention: independent nongovernmental technical evaluation.
Transluce, an independent AI research organization, provides a useful example. Its work demonstrates how outside technical organizations can independently investigate AI behavior and use AI-enabled evaluation methods at a scale that would be difficult to achieve manually.
That work is closely related to the assurance structure proposed here, but it is not the same as formal independent assurance. A mature assurance system would additionally require recognized criteria, defined levels and subjects of assurance, provider qualification, independence requirements, quality management, external oversight or accreditation, standardized reporting and accountability for the resulting assurance conclusion.
About This Project
I approach AI assurance not as an artificial-intelligence expert, but as an AI user and a certified public accountant with extensive professional experience in independent assurance over financial reporting and regulatory compliance.
AI presents technical problems far beyond the expertise of the accounting profession. But it also presents a familiar institutional problem: people are increasingly being asked to rely on complex and consequential systems whose underlying operation and evidence they cannot personally verify.
The proposal does not suggest that accountants should control AI assurance. A credible AI-assurance ecosystem would need to be multidisciplinary, drawing on expertise in artificial intelligence, computer science, cybersecurity, statistics, engineering, relevant subject-matter domains and established assurance disciplines. Competence, independence, evidence, quality management, accountability and effective external oversight should determine who is qualified to perform the work.
A note on AI collaboration
The article itself was developed with extensive assistance from OpenAI's ChatGPT for research, testing and challenging arguments, examining contrary views, drafting and editing. I remain the author and take responsibility for its thesis, judgments and conclusions.
That experience is itself relevant to the proposal: AI can make extraordinary research and analytical capability accessible, while users still lack the kind of independent assurance framework that could tell them how much reliance particular AI-assisted work deserves.