Sandeep Shilawat is a renowned tech innovator, thought leader and strategic advisor in U.S. federal markets.
getty
For years, the AI industry has asked the wrong question: Is the model aligned? But the time for that question has long passed with the advent of serious geopolitical conflicts involving AI.
As AI moves from assistant to executor, it's helping people make decisions in everything from wars and cybersecurity to benefits processing, procurement, logistics and customer operations. The most important question now is, can we trust these systems when conditions are messy, adversarial and fast-moving?
That's the real essence of investigating trustworthy AI.
Too often, trustworthy AI is described in soft terms: ethical, responsible, safe or fair. I've even used these terms for years. Those ideas matter, but they're incomplete. In practice, trustworthy AI isn't a slogan or a static property of a model. It's a property of a system that must be continuously produced, measured and enforced.
I'm defining a trustworthy AI system as one whose behavior can be observed over time, tested under pressure, evaluated independently and constrained when risk crosses a threshold. Trust isn't something we declare. It's something we engineer, and it needs to be earned over time.
This distinction matters and is even critical, because AI is now entering the operational core of institutions and nations. When a model drafts marketing copy, failure is annoying. When it helps guide cyber defense, adjudicate services or influence mission-critical decisions, failure becomes operationally significant. When a model makes an error on data or logic leading to the deaths of hundreds of people, “usually safe” clearly isn't safe enough.
That's where many current AI governance approaches break down. We need a new perspective.
Today, most organizations rely on a familiar mix of model tuning, static red teaming, strategy and policy documents and point-in-time approvals. These controls were inherited from traditional governance methods, where systems were largely deterministic and their behavior could be inspected more cleanly. Large language models (LLMs) and generative AI don't behave that way. They're probabilistic, context-sensitive and highly vulnerable to framing effects across interaction history.
In other words, the same model can behave differently depending on how it's engaged, what sequence of prompts it's seen, what authority it believes it's operating under and how pressure accumulates over time.
That's why so many AI safety programs create a false sense of assurance. They test snapshots while the real risk lives in trajectories.
This is also why the distinction between alignment and enforcement matters so much.
Alignment attempts to shape model behavior statistically. Techniques such as reinforcement learning from human feedback encourage the model toward preferred responses. That can improve behavior, but it doesn't guarantee control. Under sufficient pressure and with use of dark patterns, adversarial creativity or multi-turn manipulation, statistical preferences can erode or even fail. This has been proven multiple times in recent years.
Enforcement is different altogether. Enforcement is external to the model and it's architectural. It determines what the system is allowed to do regardless of what the model “wants” to say, feels or infers.
That's the difference between hoping a system behaves and ensuring that it can't exceed defined boundaries—a clear distinction between deterministic and probabilistic approaches.
Executives should think about this the same way cybersecurity evolved. We no longer trust users, devices or network locations simply because they appear legitimate. We verify continuously, limit privileges and assume compromise is possible. Why should anyone trust models? Generative AI requires a similar shift: zero trust for intelligence.
That means never trusting a prompt simply because it looks benign, never trusting accumulated context simply because the prior turns seemed harmless and never trusting model output simply because it sounds fluent, coherent or confident. Never trust and always verify.
Some of the most dangerous failures in AI don't show up as spectacular jailbreaks. They emerge gradually. A model resists a direct malicious prompt but begins to drift under multi-turn escalation. When authority boundaries blur, personas override policy, hidden logic leaks and biased framing accumulates while the output still sounds polished and plausible. The AI model output is no longer trustworthy.
That's the dangerous middle ground many organizations aren't measuring. So, how do we solve it?
No amount of policy or alignment will carry the burden of trust. Trust has to be engineered in the AI system.
Trustworthy AI requires assurance, which requires continuous adversarial testing, quantitative risk scoring, independent evaluation of outputs and runtime controls that can intervene when behavior crosses a defined threshold. In mature systems, this can include something like AI judges, behavioral monitors, context resets, action gating and deterministic guardrails that stop unsafe execution before it propagates. Over a period of time, you'll earn trust.
The shift is subtle but has a profound impact. We go from evaluating whether a model is good to proving whether a system remains under control.
That's the future of trustworthy AI.
Capability is becoming widely available, while trust is hard to come by. In the next phase of AI adoption, the competitive advantage won't belong to the organization with the most impressive model demo but to the one that can show, in real time, that its AI remains observable, governable and enforceable under pressure.
Trust will become the competitive advantage for AI firms. That's what trustworthy AI should mean now.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。