AI is already creeping into QHSE workflows.

Some professionals use it quietly to summarise documents. Others test it for brainstorming or rewriting procedures. A few are openly sceptical - and for good reason.

The problem isn’t whether AI belongs in QHSE. It’s knowing exactly where it helps, where it doesn’t, and where it becomes dangerous.

This article draws that line clearly, based on current, real-world AI capabilities - not future promises.

First: What AI Is Actually Good At Today

Let’s start with where AI reliably adds value right now, when used correctly.

1. Scanning Large Volumes of Documentation Consistently

AI does not get tired, bored, or rushed. It can:

• Review hundreds of pages without attention drop-off

• Apply the same lens across every document

• Surface items humans may skip under time pressure

This makes AI particularly strong as a first-pass reviewer.

2. Highlighting Gaps, Omissions, and Inconsistencies

Well-configured AI is effective at identifying:

• Missing control measures

• Risks mentioned without mitigations

• Inconsistencies across RAMS, risk assessments, and policies

Crucially, this is about absence detection, not judgement.

Humans are good at evaluating risks. AI is good at pointing out where something might be missing.

3. Reducing Cognitive Load for Reviewers

By pre-flagging areas of concern, AI allows QHSE professionals to:

• Focus attention where it matters most

• Spend less time on low-risk, repetitive sections

• Apply judgement with more mental bandwidth

This aligns directly with cognitive load research: reducing review burden improves decision quality.

4. Acting as a “Second Pair of Eyes”

AI is most effective when positioned as:

“What might I have missed?”

Not:

“Is this safe?”

Used this way, AI helps counter:

• Overfamiliarity

• Confirmation bias

• Review fatigue

Now the Important Part: What AI Cannot Do Today

This is where expectations must be managed carefully.

1. AI Cannot Make Safety Judgements

AI does not understand:

• Operational context

• Site-specific constraints

• Real-world feasibility of controls

It cannot decide whether a risk is “acceptable”, “tolerable”, or “reasonably practicable”. That responsibility must remain with competent professionals.

2. AI Cannot Replace Professional Accountability

In QHSE, accountability is explicit and personal. AI cannot:

• Take legal responsibility

• Defend decisions during enforcement

• Explain trade-offs made under real constraints

Regulators expect named, competent persons, not automated decisions.

3. Generic AI Cannot Be Trusted Without Constraints

General-purpose tools (e.g. ChatGPT-style systems) are:

• Optimised for fluent output

• Sensitive to prompting

• Capable of confident errors

In safety-critical work, plausible but wrong is a serious risk. Without domain constraints and traceability, generic AI should not be used for core risk identification.

4. AI Cannot Understand “What Matters Most” Without Guidance

AI does not inherently know:

• Which hazards are high-severity but low-frequency

• Which controls are legally mandatory vs best practice

• Which omissions are critical vs administrative

Those priorities must be designed into the system or provided by humans.

The Right Mental Model for AI in QHSE

The safest and most effective framing is simple:

AI extends human attention.
Humans retain judgement and accountability.

When AI is used to:

• Flag

• Surface

• Compare

• Highlight

…and humans are responsible for:

• Deciding

• Approving

• Signing off

You get the benefits without introducing silent failure modes. This aligns closely with emerging ISO guidance on human oversight of AI systems.

What Good AI Use in QHSE Looks Like Today

In practice, effective teams use AI to:

• Review documents before human approval

• Identify areas needing deeper scrutiny

• Support consistency across large document sets

• Reduce reliance on memory and pattern familiarity

They do not use AI to:

• Auto-approve risk assessments

• Generate controls without review

• Replace competence or training

The Real Question QHSE Leaders Should Ask

It’s not:

“Should we use AI?”

It’s:

“Where does human judgement fail under pressure — and how do we support it safely?”

Used thoughtfully, AI can reduce blind spots. Used carelessly, it can hide them.