AI in the Asbestos Sector: The Genie is Out of the Bottle?

AI can review reports, draft plans and retrieve guidance in seconds—but in a safety-critical industry, an answer that is almost right can still put people at risk.

The question is not whether AI has a place in the asbestos sector. It already does. The real question is where its assistance should end and professional judgement must begin.

Like many people in our industry, I spend a significant part of my working day writing documents, responding to emails and reviewing information. AI has helped with many of those tasks—and it has even assisted me in developing this blog, note the use of the “em-dashes (—)” retained for transparency.

Useful does not mean infallible

Upload an asbestos report and ask AI to identify errors, inconsistencies or omissions, and it may produce a useful review. But what is that review based on—and how do we know it is correct?

AI can draw on information from the internet, but it may also generate an answer without making the basis of that answer clear. Some platforms provide links or citations, but those sources still need to be opened and checked. Even the platforms themselves warn that their output can be incorrect.

The more often we receive plausible answers, the easier it becomes to accept them without verifying the facts. In asbestos work, that habit can have serious consequences.

A result may look “near enough”, and a minor error may not seem significant. Yet terminology matters. If someone asks about an “asbestos coating”, AI might interpret that as a decorative textured coating rather than sprayed asbestos and offer advice for an entirely different risk profile.

The same concern applies when AI is asked to create a Plan of Work for asbestos remediation or an Asbestos Management Plan. Is the output based on current, approved information? Does it address every relevant requirement? A user who is not fully conversant with the subject may not know what is missing. Asking AI to check its own work is not independent verification.

I recently used an AI platform to help develop a fairly complex piece of software. The task involved multiple questions, deeper explanations and cross-references to HSE and industry guidance. After several hours, the platform abruptly produced a detailed recipe for apple pie. Thankfully, I had saved my progress and could restart. The incident was amusing and the recipe was not bad, although half to one tablespoon of cinnamon seemed excessive but it also showed how quickly an AI conversation can go off course.

Assisted Information, not artificial authority

I support using AI where it is appropriate and understood. It is not there to replace people; it is there to assist them. Perhaps, in this context, AI is better thought of as Assisted Information rather than Artificial Intelligence.

AI responds to the question or instruction it is given. Ask it for a method statement without describing the work in sufficient detail, and the result may be very different from what you need.

As a deliberately unrelated test, I asked AI how a brain aneurysm might be treated. I was less interested in the medical detail than in whether the platform would answer such a specialist question. Despite my having no medical qualifications beyond first aid, it provided a detailed response with very little prompting. Information may be easy to obtain; that does not make the reader competent to apply it.

What the first answer leaves out

I also asked AI to write a Plan of Work for removing an asbestos-containing gasket. The first answer explained a broad process, precautions, decontamination and waste disposal. It appeared reasonable—but the detail was incomplete.

Only after I specified that the work would be carried out in the UK did the platform suggest it would most likely be notifiable non-licensed work. Further prompts were needed before it discussed notification and medical requirements. It recommended a P3 respirator,  but failed to mentioned face-fit testing until ask the specific question. The same pattern applied to transporting hazardous waste, insurance and training.

This is where human intelligence and professional competence must take over. AI answered the questions I asked; it did not reliably identify every question that needed to be asked. Had I not already understood the requirements, I might not have recognised the gaps.

A sensible rule for using AI

·        Treat AI as an assistant, not an authority.

·        Give it enough context to understand the task.

·        Check its output against current, reliable guidance.

·        Make sure a competent person reviews anything that could affect health, safety or compliance.

Like any tool, AI can be useful in skilled hands but damaging when it is relied upon without sufficient knowledge or scrutiny. Use it cautiously, appropriately and with care.

I’m reminded of an episode of The Twilight Zone, “The Man in the Bottle” (Season 2, Episode 2): where a shop keeper rubs a bottle and is granted 4 wishes by a Genie, it goes terribly wrong, the moral of the story? Be careful what you ask for, and make sure you are very, very clear about how you ask for it.

Post Script – I bow to my robot masters, just edging my bets with that one!!