An AI can draft your JHA. It cannot do it.
A language model writes a job hazard analysis from generic patterns, not from your floor, so it will omit the site-specific hazard a walkdown would catch and dress the omission in confident, standard-sounding prose.
Ask a large language model to write a job hazard analysis for grinding castings, and in about ten seconds you get something that looks right: task steps, hazard descriptions, a clean controls column, even the hierarchy of controls in the correct order. It reads like a JHA a competent safety professional would produce. That fluency is exactly the problem.
What a JHA is actually for
OSHA’s guidance, OSHA 3071, Job Hazard Analysis, defines the technique as one that “focuses on job tasks as a way to identify hazards before they occur,” and that “focuses on the relationship between the worker, the task, the tools, and the work environment.” Note the word “relationship.” A JHA is not a document about a job in the abstract. It is an analysis of one specific job, in one specific place, with the specific tools and people who do it.
That is why OSHA’s very first instruction under “Where do I begin?” is “Involve your employees,” and the reason it gives is blunt: “they have a unique understanding of the job, and this knowledge is invaluable for finding hazards.” The document tells you to watch the employee perform the job, list each step as the worker takes it, and include the employee in all phases of the analysis. The value is in the watching and the asking. The paper is a byproduct.
ISO 45001:2018 makes the same point as a requirement rather than a suggestion. Clause 5.4, consultation and participation of workers, obligates organizations to involve workers, including non-managerial workers, in identifying hazards and assessing risks. Consultation means seeking their views before deciding. Participation means involving them in the deciding. Neither can be performed by a model that has never been to your site.
Where the model goes wrong, and why you will not notice
A language model generates text from patterns across millions of documents. Asked for a JHA, it produces the statistically typical hazards for that task type. For grinding, it will reliably surface flying particles, laceration from burrs, and back strain from lifting, because those are the standard entries. What it cannot surface is the hazard that exists only at your site: the pinch point created by where your fixture actually sits, the exhaust ventilation that was capped during a remodel, the fact that this bay floods when it rains, the near-miss your second-shift crew talks about but never reported.
OSHA anticipated this failure in plain language. The booklet warns that its own sample procedures “are for illustration only and do not necessarily include all the steps, hazards, and protections that apply to your industry.” A generic hazard list is a starting prompt for observation, not a substitute for it. The model gives you the generic list with none of the caveats and all of the confidence.
This is textbook automation bias: the documented tendency to over-rely on an automated system’s output and to skip the vigilant verification you would apply to a human’s. The effect worsens under time pressure and cognitive load, which describes safety documentation work exactly. Fluency reads as competence. A rough, hand-scrawled JHA that a crew openly distrusts is safer than a polished one they assume was checked, because distrust triggers verification and polish suppresses it. A plausible, authoritative-looking procedure that is wrong or incomplete is more dangerous than an obviously rough one, precisely because no one re-checks it.
The honest version of the upside
None of this means avoid the tools. A model is a genuinely useful drafting assistant: it formats consistently, it remembers the controls hierarchy, it turns a walkdown’s messy notes into a clean structure, and it can prompt a novice analyst toward hazard categories they might forget. Used that way, it lowers the friction of producing documentation and frees time for the floor work that actually matters. The failure mode is singular and specific: treating the output as the analysis instead of as a first draft to be verified against the physical job.
Before you invest
Pick any AI-drafted JHA before it is used on a real task and ask: which line in this document could only have been written by someone who watched this job performed at this site? If the honest answer is "none," you are holding a template, not an analysis. Send someone to the floor, walk the task with the crew who does it, and let their observations, not the model's fluency, decide what the JHA says.
The tools will keep getting more fluent. That raises the stakes rather than lowering them, because the gap between how authoritative a JHA looks and how grounded it actually is will only widen. The discipline that protects you is old and cheap: the analysis happens where the work happens, with the people who do it. Everything else is formatting.