The risk with AI generated JSAs is that workers accept a well written list of hazards without actually thinking about the job in front of them, which defeats the purpose of the JSA. A neat, complete looking list is persuasive, and when it appears before the crew has discussed the job it tends to become the answer rather than a starting point, with people signing it because nothing on it looks wrong rather than because they checked it against what they can see.
The value of a JSA comes from the crew walking through the task and spotting what is different about today. Generic hazards like dropped objects or pinch points are useful reminders, but the incidents that JSAs are meant to prevent usually come from specifics, such as a scaffold that was modified overnight, another crew working nearby, weather that has changed or a tool that is not the one normally used, and those details only come from people looking at the actual worksite.
A better use of AI is to support that conversation rather than replace it, for example by asking the crew to identify hazards first and then showing what they might have missed, or by pulling in lessons from similar past jobs and incidents. Order matters here, since showing the AI suggestions only after the crew has recorded its own hazards keeps the thinking with the people doing the work, and it turns the suggestions into a check that can prompt a useful discussion rather than a list to accept.
The worker still owns the analysis, and the AI acts more like an experienced colleague checking their work than a form that fills itself in. The JSA record should show which hazards the crew identified and which were added after a suggestion, which gives supervisors a way to see whether crews are thinking for themselves and gives the safety team useful information about which kinds of hazards are commonly missed and might need more attention in training.
Rolling this out works best when supervisors keep doing what good supervisors already do, which is asking the crew to talk through the job and probing anything that looks copied or generic. If JSAs start looking identical from one day to the next, or if crews accept every suggestion without changes, that is a sign the tool is replacing thinking rather than supporting it, and it is worth revisiting how and when suggestions are shown.
