Offload at your peril
Artificial intelligence allows us to outsource thinking, in order to focus on the important stuff. But what impact might that have on health and safety?
Ben Hargreaves, head of content
Offload at your peril
AI allows us to stop thinking to focus on the important stuff. But what impact might that have on health and safety?
Ben Hargreaves, head of content
It would be a major irony if the race to exploit artificial intelligence ended up making us less good at doing things we already know how to do.
Some engineers, however, are warning that use of AI could risk degrading hard-won skills and know-how. If that deterioration takes place in an area where it’s vital not to cut corners – such as health and safety – it might not just be skills that are at risk. People could be too.
That’s the view of Richard Waine. Waine is a chartered engineer researching AI safety and risk management in infrastructure, and the founder of consultancy Emlyn Square. Emlyn Square has been assessing the risks and opportunities of the ‘cognitive offloading’ enabled by AI when it comes to health and safety.
Cognitive offloading is not unique to AI usage, explains Waine. If you’ve ever used a calculator or entered a location into sat nav, you’ve offloaded the mental effort of doing arithmetic or learning a route to technology. The difference is that AI is being used to do increasingly sophisticated tasks and help make complicated decisions. The ease with which it delivers answers to problems can be seductive. When it comes to maintaining an effective health and safety regime, this convenience has the potential to cause harm.
Waine explains that we are “all doing cognitive work in the knowledge economy in moving information around. We are using information to think about problems, making connections, judgements and decisions, and acting on them. When we offload that work, we transfer it to other tools and systems. Using GPS is a really good example.”
But cognitive offloading to AI is slightly different. “We are offloading deeper cognitive work and skills – things like analysis, synthesis, reasoning, and critical thinking, and we can ask AI for advice. In automated, or agentic systems, we can ask AI to make decisions on our behalf.”
The potential negative effects include not only a reduction in critical thinking [see box below], but also over-reliance on AI systems, or even ‘cognitive surrender’, in which critical thinking and decision-making is completely deferred to AI without oversight. Memory could also easily be compromised. (The International AI Safety Report 2026 says AI safety risks include deskilling and poorer decision-making.)
Of course, AI is designed to lighten the load. That’s attractive when there’s too much going on. It will even flatter you as it helps out. ChatGPT congratulates you for asking a great question and suggests a range of options. “AI can be overly helpful sometimes,” says Waine. “You'll ask it to do one thing, and it will look two or three steps ahead of that task for you.
“So, it's designed to take more off you if you let it. Some of the AI tools are designed to make you feel good about your interaction with the technology and what you've done or what you've suggested.”
Why think about it?
Emlyn Square has been working with the Health and Safety Executive Science Division on a project looking at cognitive offloading in the context of physical infrastructure, where accidents have the potential for serious consequences. As we all know, operators of infrastructure have responsibility for controlling risks using management systems in which people are responsible for reducing risks to health and safety as far as is reasonably practicable.
Historically this work has been done by people. But it could now potentially be offloaded to AI by individuals or organisations. This creates a potential link between process safety and AI safety risks such as deskilling.
As an example, people could suffer mental blank spots in the event of an emergency because they are so used to deferring decisions to artificial intelligence. Or as Waine puts it, using AI means “you are losing the opportunity to practice. If you use Google Maps for navigation, your map reading and navigation skills decline. You have weaker mental models.” With less practice, “we lose our familiarity with how things work. And that becomes important when we're under pressure, or there's uncertainty, because we will typically fall back on those mental models to make decisions.”
Waine adds: “When things don't go as we expect them to – an emergency or incident is a good example – we've potentially lost those mental models and our judgment has degraded. If you take away the AI from that situation – let's say you must respond in analogue only – can you still do it?”
These concerns have surfaced just as usage of AI is rising in health and safety for tasks such as writing risk assessments or analysing risk using video technology, says Waine. “AI is already influencing health and safety decision-making.”
Such usage may not be officially sanctioned by the organisation. Emlyn Square’s research highlights ‘shadow’ or ‘grey’ AI where the technology is used outside the structures of organisational IT provision. This use of AI on the margins may not be visible or controlled (a good example is employing a consumer tool such as ChatGPT to carry out work tasks) but still make a big contribution to health and safety-related decision-making.
All of this means a collaborative effort by the energy and water sectors to develop further understanding and frameworks for the use of AI in health and safety is now needed, says Waine.
“We need to be pragmatic about this with the pace of AI development. One of the contributing factors to the use of shadow AI is that the frontier models are so much better in comparison to those models that are just six months old. So, when we're thinking about governance frameworks and standards, they must remain relevant and practical to control risks as they emerge.”
It’s not all bad. Waine stresses the positive impact of cognitive offloading in improving productivity and efficiency. He says well-informed AI users will learn how much and where to cognitively offload and how to scrutinise the technology. “The challenge is how do we create the conditions for appropriate cognitive offloading,” Waine says.
Get that right, and AI will help make utilities safer. For tasks that are repetitive or monotonous, automation has the potential to free up engineers to analyse more complex risks.
But it’s vital engineers maintain responsibility for what AI is doing. In August, Ofgem released high-level findings from its AI Reg Lab, which included a statement that in all cases, people should remain accountable for outcomes.
Until this changes, people will remain a critical part of decision-making. “They will need the skills and decision-making competence to operate in an AI-augmented environment,” says Waine.
The security threat
Along with the potential impact of cognitive offloading on health and safety, AI raises fresh concerns when it comes to the cyber security of critical national infrastructure (CNI).
Cyber security remains a key concern for operators of CNI, as the high-profile hack of an unnamed peaker plant in August and a surge in cyber-attacks on water and wastewater plants in the US earlier in the summer demonstrated.
These were just a couple of the more well-documented incidents. The chief executive of the National Cyber Security Centre (NCSC), Dr Richard Horne, recently said more than 200 cyber incidents affecting the UK’s critical national infrastructure and its supporting ecosystem were managed by the NCSC in the year to May 2026, with around 75% of those believed to be linked to state actors. Cyber security consultancy Bridewell also found in a study published this year that 93% of CNI organisations experienced a cyber-attack in the past 12 months.
Number of cyber attack against critical infrastructure managed by the NCSC in the year to May
Proportion of these attacks believed to be linked to state actors
Proportion of critical infrastructure companies that have experienced a cyber attack in the past 12 months
There are many more cyber-attacks on CNI taking place than those that make the news, points out Graeme Stewart, director of public sector sales at Check Point Software Technologies (he believes “breach fatigue” from the sheer number of attacks is responsible). When an incident does make headlines, such as the attack on carmaker Jaguar Land Rover, it’s because it’s a big deal. “There was a report that said the JLR attack was so big, it actually had a fundamental effect on the GDP of the country. I've been involved in cyber for nearly 30 years, and the idea that a cyber-attack would have a material, measurable effect on UK GDP is mind-blowing.
“My fear is that at some point there will be a successful attack on something in the CNI.” A lack of energy supplies could lead to a meltdown of law and order within 24 hours, Stewart suggests.
He says AI can be used as a tool by hackers to enable them to access areas they might otherwise struggle to get into. It may also help those with little hacking proficiency to become effective operators, while ratcheting up the scale of attacks. “It is speeding up the creation of attacks, but it's also speeding up the orchestration of the attacks.” (In the US, the expression ‘quarterbacking’ is now used to refer to a single coordinator or lead strategist managing multiple moving parts using AI.)
Stewart says: “The orchestration of a cyber-attack can be controlled using these AI tools. And frankly, someone with a little bit of knowledge can suddenly become exponentially more dangerous.
“The flipside to this is that the cybersecurity industry is using AI to rapidly improve the quality of the cybersecurity tools we deploy, as well as providing guardrails and controls for AI. We’ve got a whole section of our technology portfolio that is just focused on AI because the bad guys are using it as well.”
Networks are well-aware of this. “We have always known it was coming down the line,” says a security expert at one electricity network. He highlights the public announcement that Anthropic’s Claude Mythos model is capable of discovering hard to detect vulnerabilities (known as ‘zero-day’ vulnerabilities, or security flaws that no one knows about) in software. This prompted Anthropic to share the technology with AWS, Microsoft, Apple and other tech giants to develop defensive capabilities earlier this year.
But the Anthropic revelation also means hostile nation states and criminals are aware of the hacking potential of systems like Mythos. “There are already China-based organisations that are delivering tools like that,” the security expert says. “That means criminals can scan all your public-facing applications to find out how many vulnerabilities there are where you are not protected.
“That is the biggest threat to most CNI organisations: the pace of discovery. They can find out vulnerabilities in a matter of hours. We will not get patches in a matter of hours. Because of the pace of discovery, it has become even more important for CNI organisations to have mature cyber governance and incident response in place.”
The first port of call for organisations that want to ensure they are as protected as possible is the NCSC site, says Graeme Stewart. The NCSC Cyber Essentials programme helps organisations protect themselves against common online threats, providing a baseline level of security to work up from.
Companies should consider all people, processes and technology that could impact on security, and what to do in the event that they are hacked. ‘Wargaming’ scenarios can help organisations identify how they would respond if the worst happens. Stewart adds that it is also a good idea to consider the cyber security credentials of the supply chain. “Your supply lines are your weakest point.”
Like Waine, he stresses the positive impact of AI. “In the right hands, it's this incredibly valuable tool that will deliver productivity gains and autonomous thinking and creativity.
“In the wrong hands, it's a weapon.”
The ‘5 Ds’ of cognitive offloading risk patterns:
- Dependency – cognitive offloading may degrade engineering judgment and capability.
- Disengagement – cognitive offloading may reduce opportunities to reorganise how attention is focused on systems.
- Drift – cognitive offloading may allow professional understanding to fall behind the AI system status.
- Data – cognitive offloading may change what information is valued by an organisation.
- Disconnection – cognitive offloading disconnects social relationships, professional collaboration, and training.
