A new study says AI is quietly eating the safety guardrails out of industrial control software.
Researchers examined 62 real-world Simulink controllers, the kind of software that runs factory equipment, vehicles, and other machinery, comparing traditional hand-coded designs against newer AI-enabled ones. They sorted the models into 8 controller types across 10 application domains and surveyed 13 practicing engineers who build this stuff for a living. The analysis found that subsystem organization, basically the scaffolding that holds a controller together, eats up 68 to 72 percent of a model's footprint in both approaches, while the actual control logic takes up very little space. AI-enabled controllers also leaned heavily on discrete dynamics and custom, user-defined abstractions that barely show up in the academic literature describing how these systems are supposed to work.
The more striking finding is what disappeared. Explicit constraint-enforcement blocks, the visible checks that stop a controller from doing something dangerous, largely vanish in AI-enabled models even though the surveyed engineers expected them to be there. That suggests safety is shifting from code a reviewer can point to and audit into whatever the training process happened to learn, which is much harder to trace or verify after the fact.
For an industry still arguing about whether a neural network belongs anywhere near a robot arm or a car's brakes, a traceability gap this size is less a footnote than the whole plot.