Generative AI in a safety system means using a language model to describe, explain or recommend - as an advisory layer over decisions made elsewhere, not as the mechanism that makes them.
A language model is an excellent explainer and a poor safety authority. It can be unreachable, it costs money per call, and it can be fluently and confidently wrong. Any one of those disqualifies it from deciding whether an alarm sounds.
Deterministic logic decides. The model is given the readings and that decision, and asked to produce a severity assessment, a plain-language cause and a recommended action for a human reader. Its output is presented as advice and never actuates anything.
'Carbon monoxide is rising steadily near the furnace and will reach the alarm limit in roughly twenty minutes; check the flue' is more actionable for a supervisor than a channel name and a number.
The model is invoked only after the deterministic layer has classified a reading as worth explaining, and its unavailability changes nothing about the alarm.
