Workers blame leaders and safety teams, not employees, when unreviewed AI safety documents cause injury
Only 18% of workers say their workplace has a formal policy defining who is accountable when AI-assisted safety content contains an error, even as generative AI takes on more of the drafting of procedures, hazard assessments and toolbox talks, according to a new survey.
The study by TRADESAFE, a U.S. supplier of lockout/tagout devices and other workplace safety equipment, polled 514 U.S. workers in manufacturing, construction, warehousing, logistics and the skilled trades in September 2026.
Among the 399 respondents with visibility into how safety documents are created, 51% said their organisation uses generative AI to help write them. Just 49% of respondents said those AI outputs are independently reviewed before use.
Herbert Post, vice president of TRADESAFE, told HRD that the lack of defined ownership is the main problem.
"Most workers can't point to a policy that says who owns an error in an AI-drafted safety document. In our study, only 18% said their workplace has one. That is the real gap," Post said.
"A lockout procedure or an emergency instruction needs a named person who signs it off and answers for it. Without that, 'the AI got it wrong' becomes the default answer, and a worker on the floor is following a procedure nobody has actually put their name to".
Leaders, safety staff see review differently
Perceptions of oversight varied sharply by role. Company leaders (74%) were far more likely to report independent review of AI-drafted documents than supervisors or managers (52%), frontline workers (36%) or safety professionals (23%).
Frontline workers were also far less likely than leaders to say they had visibility into how safety documents are produced (31% versus 76%).
"Leaders may believe that a qualified check is done that workers do not see," Post said.
The survey suggests errors are already reaching workers. More than a third (36%) said they had personally seen an AI-generated safety document with an outdated requirement, a vague instruction or a missing hazard step.
One in five (20%) reported seeing an adverse outcome linked to such an error. Near misses were the most common (11%), followed by minor injuries (7%), failed inspections (6%), equipment or property damage (5%) and serious injuries (3%).
About a third (34%) said AI is being adopted for safety work faster than their company can properly oversee it.
Blame moves up the chain
In a randomised scenario, respondents were shown a safety procedure containing a mistake. The procedure was presented as written by a person, drafted by AI and then reviewed, or drafted by AI with no review.
When the unreviewed AI draft led to a serious injury, 56% blamed company leadership or the safety department (28% each). Only nine per cent blamed the employee who handled the document. The combined share for leadership and the safety department was 49% for the human-written version and 39% for the reviewed AI draft.
Nearly two-thirds (64%) agreed that blaming AI for a mistake usually means a company's oversight is not clear enough.
Trust also depended on review. Sixty-three per cent said they would trust a safety policy less if they knew it was partly AI-written, while 78% said review by a qualified person increases trust.
"What workers distrust is the missing review. In our study, 78% said that they trust a safety document more when a qualified person reviews it, regardless if AI helps in drafting it. That should alleviate employers' concerns," Post said.
"Using AI to draft safety documents faster is acceptable, as long as a qualified person reviews it, and it is signed off by someone who is knowledgeable about the task and site before it's posted to the shop floor. Reviewing the safety document makes it acceptable and shows the employer stands behind what it gives to the worker."
