This is the first article in the Securing the Future of Industrial Energy Production series.

Industrial facilities do not suffer from a shortage of complexity. They suffer from a shortage of visibility: of knowing when something will break, why a signal just glitched, or whether a seemingly minor event is an indicator of catastrophic failure to come. This visibility gap is exactly the space AI promises to fill. But as with any system that offers extraordinary power, the promise comes laced with risk.

If you’re a CTO, VP of Engineering, Director of OT Security, or Head of Plant IT responsible for safety and security across an energy, manufacturing, or infrastructure operation this article is for you. Especially if you’re stretched thin, leading a hybrid team, and trying to modernize legacy systems without introducing new threats. The rise of AI in industrial environments isn’t just hype, it’s directly impacting the systems you’re tasked with keeping safe and secure.

The evangelists say AI will revolutionize industrial safety. The skeptics say it will inject new failure modes into systems that already teeter on the edge of acceptable risk.

Both are right.

The Safety Systems We’ve Come to Trust

Industrial safety has historically been an exercise in deterministic engineering. We build physical redundancies, install PLCs with hard-coded fail states, isolate critical components behind air gaps, and layer safety interlocks like armor. In nuclear plants, for instance, we model fault trees down to improbably rare sequences of valve and sensor failures. These models are based on physics, human behavior, and known failure data not statistical guesses.

There’s comfort in the predictability. If a temperature sensor fails, a backup kicks in. If a motor overheats, it’s shut down. These systems might not be smart, but they are legible and legibility is the currency of trust in high-consequence environments.

For the industrial decision-maker managing legacy control systems now being tethered to cloud services or remote analytics, these traditional systems offer a false sense of finality. You know you can’t just patch AI onto safety like an upgrade. You need systems designed with safety and security in mind from the start.

Enter AI: The Prodigal Data Whisperer

AI excels at pattern recognition in noisy data. Predictive maintenance tools now detect failure precursors in vibration signatures that human engineers never would’ve noticed. AI-driven computer vision can spot microfractures in gas pipelines or track corrosion on structural elements across years of inspection images.

These are not small wins. They prevent real-world injuries, fires, and outages.

For leaders running industrial production systems, this opens up operational possibilities. Imagine replacing fixed maintenance schedules with dynamic, data-driven interventions. Or deploying drones with onboard AI to inspect high-risk assets that previously required scaffolding or personnel exposure.

AI enables a move from reactive to proactive safety. Instead of responding to alarms, operators are increasingly being nudged by AI systems that say, “This bearing has an 87% chance of failing within 96 hours.” For aging electrical infrastructure, this predictive capability borders on miraculous.

In some cases, AI even navigates environments no human safely can. Drones powered by onboard neural networks inspect high-voltage lines, while quadruped robots monitor radioactive zones. AI gives us not just data, but reach.

So, what’s the problem?

The Problem Is That AI Doesn’t Think. It Predicts.

And prediction is not the same as understanding.

AI systems, especially those based on deep learning, do not reason about cause and effect. They interpolate from the data they’ve seen. If that data is flawed, poisoned, or just incomplete, the results can be confidently wrong in ways that defy operator intuition.

More concerning is the black-box nature of many AI systems. When a safety system says, “shut the reactor down,” we want to know why. But try asking a convolutional neural network to explain its classification of a coolant leak.

Then there’s the risk of model drift, where models gradually lose accuracy as the environment changes, or automation bias, where operators begin to trust the AI so completely that they ignore red flags. In the worst-case scenario, you get a system that fails silently, and a human that assumes it hasn’t.

If you’re leading a facility with minimal in-house software security depth, this isn’t just an IT problem. It’s a safety problem. And one that can undermine your entire risk posture if left unexamined.

Safety Doesn’t Scale With Complexity. It Scales With Predictability

What makes AI powerful also makes it fragile. It generalizes from data but rare events, especially the black swans of industrial safety, often lie outside that data. Consider an AI that’s never seen a certain combination of process pressures and voltages. How will it respond? Will it extrapolate safely, or veer into an unsafe state?

We’ve already seen glimpses of this risk. In 2021, an autonomous inspection drone in an oil facility misclassified a critical pipe junction due to lighting artifacts and failed to raise an alert. In another case, an AI model used to predict transformer lifespan underestimated degradation due to unseen contamination patterns – patterns only a veteran technician noticed during routine inspection.

These are not failures of hardware. They’re failures of epistemology.

For the safety-minded executive, this means AI isn’t just another tool in the toolbox. It’s a tool that changes how the entire toolbox works. And it demands a new safety mindset – one rooted in both statistical reasoning and engineering rigor.

Engineering for Safe AI: A New Safety Discipline

To integrate AI into safety-critical systems, we need new tools and practices, not just better algorithms. Some of the most promising include:

• AI Explainability Techniques: Local interpretable models (e.g., LIME, SHAP) that help explain why an AI made a decision, even if the model itself is opaque.
• Hybrid Systems: Where deterministic safety logic is layered underneath or around AI systems, preventing unsafe outputs from propagating directly into the control loop.
• Digital Twins with AI Sandboxing: Allowing AI to make decisions in a mirrored environment first, observing how they behave under simulated stress before deployment.
• Red Teaming AI Models: Subjecting AI systems to adversarial conditions, including model inversion and poisoning, to test robustness.

If you’re running production systems that can’t afford downtime or safety incidents, you need to evaluate where AI fits into your architecture and, more importantly, where it doesn’t. This is especially critical when AI interacts with operational technologies (OT), where failure can mean physical damage, environmental risk, or human harm.

And none of this will matter if we don’t also train the humans. AI doesn’t eliminate the need for expert operators; it raises the bar for what those operators need to understand. In many ways, AI demands more domain expertise, not less.

So, Is AI a Game Changer or a New Risk?

Both.

It’s a powerful augmentation to human safety capability, but only when it’s treated as a fallible partner, not an oracle.

The great danger is not that AI will cause catastrophic failure. The great danger is that we will assume it won’t. And in complex systems, misplaced confidence is the root of most disasters.

If we’re to deploy AI in the crucible of industrial safety, we must hold it to a higher standard than accuracy. We must demand explainability, resilience, and integration with the disciplines that have kept our factories, power plants, and pipelines safe for the last century.

If you’re leading industrial safety and security efforts while simultaneously modernizing your tech stack, the integration of AI will be one of your most critical – and risky – strategic decisions. It’s not just a question of can we use AI, but how and where it can safely operate.

We need AI to be safety-aware. But more urgently, we need to be AI-aware.

If This Resonates With You, Here’s What To Do Next

If you’re a CEO, COO or senior leader and are interested in modernizing safety- and security-critical industrial systems, I invite you to:

👉 Connect with me on LinkedIn
👉 Look for more articles in this series coming to LinkedIn or maddphysics.com
👉 Visit laurellye.com or schedule a quick call

I publish tactical and strategic insights on securing industrial systems, integrating AI safely, and building resilient architectures that don’t just survive modernization, they lead it.