Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

What AI won’t do: Why assurance must go beyond compliance

The future of AI depends on systems that can earn trust.

[Photo: Getty Images]

Profile photo of Majeed Javdani

President · International Digitalization Forum

In an age where artificial intelligence is redefining the boundaries of possibility, the spotlight has increasingly turned to trust. We ask: Can we trust these systems? Will they make fair decisions? Can they be manipulated? And what, exactly, is the basis of that trust?

As technologists and policy leaders search for ways to answer these questions, assurance—the process of justifying confidence in a system’s behavior—has emerged as a cornerstone. It is not a new idea; assurance has long been central to fields such as aerospace and cryptography. But in AI, it takes on a unique complexity. These systems learn, adapt, and behave in ways that challenge conventional verification. It’s no longer enough to show what a system does. Increasingly, we must prove what it does not and cannot do.

ASSURANCE MATTERS

The U.S. National Institute of Standards and Technology defines assurance as the grounds for justified confidence that a claim or set of claims has been or will be achieved. While this may sound abstract, its implications are very concrete—especially in the context of AI. Assurance is not a checklist or a stamp of compliance. It’s a body of reasoned evidence that supports claims about how an AI system will behave under real-world conditions—and just as importantly, how it will not behave.

Too often, especially in regulatory environments, assurance is confused with compliance. But compliance is merely a starting point. It’s a form of what we might call “veneer security”—a surface-level signal that rules have been followed. The danger is in assuming that following rules automatically translates to trustworthiness. It doesn’t. Compliance-based assessments may play a role in informing assurance, but they cannot replace it. A compliant system can still be fundamentally unsafe, unpredictable, or vulnerable to abuse.

The stakes are even higher when it comes to AI, because its behavior is emergent—it arises from training data, optimization functions, model architecture, and system design. This makes traditional forms of assurance, like functional testing or behavioral observation, insufficient. We can no longer rely solely on evidence of what the system does when operating “as intended.” Instead, we must provide robust evidence of what the system cannot be made to do—even when pushed to its limits, even under adversarial conditions.

NON-BEHAVIOR—WHAT AI SYSTEMS WON’T DO

This is where the concept of non-behavior becomes critical. Non-behavior refers to the things an AI system does not, will not, or cannot be forced to do. It’s not about performance boundaries or system limitations—it’s about provable absence. This distinction matters. For example, it’s not enough to show that a facial recognition system usually avoids bias. Assurance must prove that the system cannot be forced into biased behavior. That bias is structurally excluded—not merely absent in the current dataset or use case. This is the core of a non-behavior judgment.

In reality, this level of assurance can’t be improvised or retrofitted. It requires intention from the start—baked into the architecture, not sprinkled on in review. Just as architects design buildings to resist collapse, AI engineers must design systems that resist misuse. This means thinking differently about development, treating assurance not as a test to pass, but as a property to design for.

Such judgments are especially vital for socially sensitive claims—non-bias, non-manipulation, non-deception. These claims carry enormous implications, both ethical and legal. If a system is marketed as unbiased, then the assurance case must prove that bias is not only unintended, but structurally impossible. That means the evidence must go deeper than performance metrics or audit reports.

SYNTHETIC ASSURANCE

Synthetic assurance is a method of building assurance not from a single test or certification, but from the ground up—through the composition of verified components at every level of the system. It is a holistic approach that sees assurance not as a final step, but as a continuous design principle. Each module, each decision function, each data pipeline contributes to the final judgment. This method ensures that trust is not an afterthought or a hopeful conclusion. It is engineered into the system.

If assurance is to be meaningful in AI, then it must embrace this synthetic model. Just as we use mathematical proofs to ensure a cryptographic algorithm cannot be broken, we must construct AI systems so that certain dangerous behaviors are mathematically or structurally impossible. This doesn’t mean every system must be provably perfect—but it does mean that critical claims, especially claims of non-behavior, must be backed by a rigorous assurance case.

It looks like a system where bias mitigation is not just a post-hoc fix, but a constraint in the optimization process. It looks like traceability in training data that links model behavior to design intent. It looks like formal methods in sensitive subsystems, continuous validation in operation, and adversarial testing that goes beyond expected use cases. It looks like teams working not just to make AI smarter, but to make its limits intelligible and its risks bounded.

In the end, assurance in AI is about accountability. It’s about being able to stand behind the system and say: We know not just what this AI can do, but what it cannot do—and here’s the evidence. That confidence cannot be borrowed from compliance, or outsourced to audit firms, or approximated through statistics alone. It must be built.

The future of AI depends on systems that can earn trust—not with marketing slogans, but with technical rigor. That future is possible. But only if we move past shallow demonstrations of compliance, and commit to the deeper, more difficult task of proving non-behavior. Only if assurance becomes not a checkbox, but a discipline.

As AI continues to shape our world, our ability to trust it will depend not on what it claims to do—but on what we can prove it cannot.

Majeed Javdani is president of the International Digitalization Forum.