Artificial intelligence is rapidly moving into the real world. It is no longer confined to research labs or recommendation engines. Today, machine learning is being used in autonomous vehicles, robotics, healthcare systems, and industrial environments. These are domains where safety is not optional. It is essential.

That shift makes the question immediate, not theoretical. Can AI be trusted in safety-critical applications?

Why This Matters Now
For many organizations, the answer is caution. Some are told to wait. Others are told that the risks are too high. At the same time, competitive pressure is increasing. AI is becoming a differentiator across industries, and delaying adoption is not always a viable option.

The tension between innovation and safety is now one of the defining challenges facing engineering teams. The choice should not be between moving forward and staying safe. The real challenge is how to do both.

Why AI Feels Different
Part of the difficulty comes from how different machine learning is from traditional software. Conventional systems follow explicit, human-written rules. Their behavior can be traced, tested, and verified.

Machine learning systems instead learn from data. Their behavior depends on the data they are trained on and the environments in which they operate. That makes them harder to predict, explain, and verify in the ways engineers are accustomed to.

Existing safety standards recognize this challenge, but often provide limited practical guidance on how to manage it in a structured and repeatable way. As a result, many organizations are left navigating uncertainty.

A Shift in Thinking: Safety Across the Pipeline
A key insight from recent work is that the risk in machine learning systems does not reside in the model alone. It exists across the entire development process.

From how a problem is defined, to how data is collected and prepared, to how models are trained, evaluated, and deployed, each step introduces opportunities for failure. Issues such as incomplete data, inconsistent labeling, or unclear system boundaries can propagate through the pipeline and emerge later as unsafe behavior.

These findings have led to a shift in thinking. Instead of focusing only on model performance, organizations are beginning to treat the entire machine learning lifecycle as a safety-critical process.

A Practical Method: ML FMEA
Machine Learning Failure Mode and Effects Analysis, or ML FMEA, builds on this shift. It adapts a well-established safety engineering method and applies it to machine learning development.

Rather than asking only whether a model performs well, ML FMEA systematically evaluates what could go wrong at each stage of the pipeline, what the impact would be, and how those risks can be mitigated. It creates a direct connection between development decisions and system-level safety outcomes.

In humanoid robotics, similar challenges emerge in different forms. Consider a robot operating in a warehouse alongside people. A perception system may fail to detect partially occluded or seated individuals, leading to unsafe proximity. In another case, a learned control policy may behave unpredictably when exposed to real-world conditions that differ from training. These issues are not simply model problems. They are the result of gaps in data, assumptions, or validation across the development pipeline. Applying a structured analysis reveals these failure modes and enables targeted mitigations, such as improving data diversity, refining training signals, or introducing runtime safety constraints.

Both examples highlight an important point. The risks are often not obvious until the system is examined holistically. A structured approach makes those risks visible earlier, when they can still be addressed.

Despite this diversity, the group was united by a shared mission: to advance the safe use of artificial intelligence in safety-critical applications. That collaboration is reflected in the method itself. It is designed to bridge disciplines, align teams, and provide a common framework for addressing one of the most complex challenges in modern engineering.

Moving Forward
Safety and innovation do not need to be in conflict. With the right methods, safety can enable the adoption of machine learning rather than limit it. The question is not whether AI can be used in safety-critical systems, but how it can be done responsibly and systematically.

More on the approach is available in the underlying work:

• Introducing the ML FMEA: A Safe Machine Learning Approach (pre-print) (published)

• ML FMEA in Action: Lessons from Applications of Machine Learning Safety (pre-print) (published)

The challenges of deploying AI safely are real, but so are the methods that address them. Organizations that have been advised against AI adoption in safety-critical contexts are encouraged to examine the available methods.

Organizations exploring safe machine learning adoption or structured approaches to risk management are encouraged to reach out to Reynolds & Moore. For those considering ML FMEA, the team is available to discuss how it can be applied to a specific system.