ISO/IEC TS 22440 Tackles AI Safety in Industrial Automation

Industrial automation increasingly relies on artificial intelligence for tasks ranging from machine vision to autonomous decision-making. Yet AI's data-driven, adaptive nature creates fundamental challenges for traditional functional safety frameworks designed around deterministic software.

A new technical specification aims to bridge that gap. ISO/IEC TS 22440 establishes principles for applying functional safety standards to AI-enabled systems in industrial settings, addressing the unique risks that emerge when machine learning takes on safety-relevant functions.

The AI safety challenge

Conventional software behaves predictably: given the same inputs, it produces the same outputs. Machine learning systems operate differently. They adapt based on training data, can produce non-deterministic results, and may behave in ways their developers didn't explicitly program.

This creates new questions for safety engineers. How do you verify a system that learns? How do you validate performance across scenarios the system hasn't encountered? What failure modes exist that don't appear in traditional software?

ISO/IEC TS 22440 addresses these questions by integrating the AI lifecycle with established functional safety practices. The specification covers hazard analysis tailored to AI systems, identification of AI-specific failure modes, and approaches to risk reduction that account for machine learning characteristics.

Industrial applications in scope

The framework applies to several automation domains where AI is gaining traction:

• Object identification and classification using computer vision for quality control or safety monitoring
• Anomaly detection through machine learning algorithms that identify equipment faults or unsafe conditions
• Predictive diagnostics that anticipate failures before they occur
• Decision-making systems where AI supports or controls safety-critical operations

Each application introduces distinct safety considerations. An AI vision system that misclassifies an object could allow defective products through. A predictive maintenance algorithm that fails to detect an impending failure could lead to equipment damage or injury.

Why it matters

Manufacturers face a practical dilemma: AI offers significant operational advantages, but deploying it in safety-critical systems without clear standards creates liability exposure and regulatory uncertainty. ISO/IEC TS 22440 provides the common framework the industry needs to move forward confidently. For automation suppliers, it establishes expectations for how AI safety should be documented and validated. For end users, it offers criteria for evaluating AI-enabled equipment. As regulatory bodies worldwide grapple with AI oversight, technical specifications like this shape the requirements that will eventually become mandatory.

Training availability

A3 members can access specialized training on the new specification through Reynolds & Moore, with a two-day course scheduled for October 14–15, 2026, at MassRobotics in Boston.

These details were first reported by Automation Watch, the news service of the Association for Advancing Automation.