Organisations deploying artificial intelligence into knowledge work are, whether they recognise it or not, running a second and largely unmeasured experiment in cognitive health. The first experiment is whether AI actually improves productivity, and the second is what sustained AI use is doing to the cognitive capacity of the people relying on it for work. The evidence on that second question is still early, but it is no longer absent in workplaces, and it deserves the same discipline applied to any other emerging risk: work from primary sources, and keep a clear line between what is established and what is still preliminary. This article examines cognitive health and safety effects from AI use.
Brain capital
In January 2026, the McKinsey Health Institute (MHI), working with the World Economic Forum’s Brain Economy Action Forum, released The Human Advantage: Stronger Brains in the Age of AI. The report introduces “brain capital” as a combined asset: brain health, meaning optimal brain functioning and the prevention or treatment of mental, neurological and substance use disorders, and brain skills, meaning the cognitive, interpersonal, self-leadership and technological literacy abilities that enable people to adapt, relate and contribute meaningfully (Coe et al. 2026).
The report’s argument is that this asset is now strategically decisive, because AI’s value depends heavily on the quality of human judgement applied on top of it. It reports employer survey data indicating that 59 per cent of employees, on average, are expected by their employers to need additional training by 2030 to meet evolving skill demands (World Economic Forum 2025, cited in Coe et al. 2026). Its warning is explicit: without deliberate investment in brain health and brain skills, organisations risk “driving up preventable costs through declining employee well-being” (Coe et al. 2026).
The report treats cognitive health as a precondition for the AI agenda, not a wellbeing matter developing beside it.
What the evidence actually shows
The discussion of “AI and your brain” has likely moved faster than the underlying science, and precision here is not pedantry. It is the difference between a defensible position and an overstated one for employers engaging people to complete AI related knowledge work.
The strongest peer-reviewed evidence to date comes from Technology, Mind, and Behavior, a journal published by the American Psychological Association (APA). A 2026 study by Baldeo, involving 1,923 online adult participants from the United States and Canada completing simulated work tasks with commercial AI tools, found that 58 per cent of participants agreed that AI “did most of the thinking” on the task, particularly for planning and sequencing work. Those participants reported reduced confidence in their own independent reasoning and a diminished sense of ownership over the ideas produced. Participants who actively modified or challenged AI outputs, rather than passively accepting them, reported greater confidence and a stronger sense of authorship (Baldeo 2026). Two qualifications matter. The study measured self-reported confidence and perceived ownership, not underlying capability, and the findings are correlational rather than causal (American Psychological Association 2026). Within those limits, the finding is still useful: passive AI reliance is associated with reduced confidence in one’s own reasoning, and active engagement with AI output is associated with retaining it.
You might like to read AI Governance in health & safety
A second, more widely cited study which likely needs a level of caution is the work by Massachusetts Institute of Technology (MIT) Media Lab. The research paper used electroencephalography (EEG) to track 54 participants writing essays either unassisted, with a search engine, or with an AI large language model. Across the main sessions, participants in the unassisted group showed the strongest and most distributed neural connectivity, AI-assisted participants showed the weakest, and AI-assisted participants also struggled to accurately quote from essays they had just produced. In a smaller follow-up session, AI users reassigned to write unassisted showed weaker neural engagement than participants who had worked unassisted throughout (Kosmyna et al. 2025). This is a widely discussed study, but it is a preprint. It has not yet cleared peer review, the sample is small (54 participants across the main sessions, only 18 in the follow-up session), the task was narrow, and a formal academic comment has flagged methodological concerns and urged more conservative interpretation of the results (Stanković et al. 2026). It is enough to justify organisational caution about how AI is used. It is not enough to support the more dramatic claims about “brain damage” that have circulated in secondary coverage.
A third, well-established line of research concerns attention itself, independent of AI. Gloria Mark and colleagues at the University of California, Irvine, have spent nearly two decades documenting the cost of interruption and context-switching on the quality of knowledge work (Mark, Gudith & Klocke 2008; Mark 2023). This matters directly to the AI question, because AI tools are frequently deployed into workdays that are already fragmented by notifications and competing demands. The judgement McKinsey’s analysis positions as the human contribution AI cannot replace (Brassey et al. 2026) is difficult to exercise without uninterrupted time to think, regardless of how good the AI is.
Together, these three strands describe a consistent picture without needing to overstate any single study: passive AI reliance is associated with reduced confidence in one’s own reasoning (peer-reviewed, correlational); heavy, uncritical AI use may also reduce the neural engagement associated with deep original thinking (preliminary, contested); and both problems are compounded when workers have no protected time to think, a problem that predates AI and is made more urgent by it. One limitation should be stated plainly: no published study yet examines cognitive offloading in Australian workplaces or in safety-relevant roles specifically, so the workplace application that follows is inference from adjacent evidence rather than direct measurement.
The cognitive health and safety dimension
For Australian workplaces, none of this sits outside existing law. Cognitive load is already a recognised hazard under the model work health and safety framework, as Safetysure’s psychosocial risk management guide sets out in detail.
The model WHS Act and model WHS Regulations, maintained by Safe Work Australia, have been adopted with local variations by the Commonwealth and every state and territory except Victoria, which regulates psychosocial hazards under its own Occupational Health and Safety (Psychological Health) Regulations 2025. The completion of that national transition is examined in Safetysure’s analysis of regulating psychological health at work. The model provisions have legal effect only as enacted in each jurisdiction, so the local instrument should always be checked, but the architecture is consistent. Under the model WHS Regulations, a psychosocial hazard is a hazard arising from, or relating to, the design or management of work, the work environment, plant, or workplace interactions or behaviours, that may cause psychological harm (r 55A). A person conducting a business or undertaking must manage psychosocial risks (r 55C), and in determining control measures must have regard to, among other matters, “the design of work, including job demands and tasks” and “the systems of work, including how work is managed, organised and supported” (r 55D(2)(c) and (d)).
The model Code of Practice on managing psychosocial hazards makes the cognitive dimension explicit. Its description of high job demands includes sustained concentration or vigilance where accuracy is required or workers are watching for infrequent events, tasks that exceed a worker’s capacity or competency, the absence of systems to prevent individual error, and rapid or repeated task-switching that makes it difficult to concentrate (Safe Work Australia 2022). Its low job demands category covers the mirror problem: long idle periods in which a worker must monitor a process and cannot do anything else. And poor organisational change management, meaning insufficient consultation and a failure to consider new hazards when planning and implementing change, is a listed hazard in its own right (Safe Work Australia 2022).
Set an AI deployment against those descriptions where a deployment that removes routine tasks and concentrates complex judgement calls into every remaining hour is a change to job demands. A deployment that moves a worker from doing the analysis to monitoring an AI doing it creates the same sustained-vigilance and idle-monitoring conditions the Code describes. Both fall squarely within the matters r 55D requires a duty holder to consider, and both engage the primary duty to provide and maintain safe systems of work under s 19(3)(c) of the model WHS Act. The consultation duty is also engaged: s 49(d) requires consultation with workers when proposing changes that may affect their health or safety, and a material redesign of how cognitive work is performed is plainly capable of being such a change. In my consulting experience, AI deployments are rarely run as though this duty applies to them. Safetysure’s research on artificial intelligence at work examines the broader regulatory picture, including the psychosocial hazards the Commonwealth Code names that the model Code does not.
There is also a longer evidence pedigree here than the current AI discussion acknowledges. Bainbridge’s “Ironies of Automation” described the core problem in 1983: automate the routine, and you leave a de-skilled human responsible for handling precisely the abnormal situations automation cannot, at the moment their skills are least practised (Bainbridge 1983). Parasuraman and Riley documented the related phenomena of automation misuse and complacency through the 1990s, largely from aviation and process control (Parasuraman & Riley 1997). Read against this literature, the MIT and APA findings look less like the discovery of a new hazard and more like early evidence that a well-documented human factors problem, previously confined to cockpits and control rooms, now applies to general knowledge work. That framing matters legally as well as scientifically: a hazard with a forty-year evidence base is difficult to characterise as unforeseeable.
For officers, the implication runs through s 27 of the model WHS Act. Due diligence includes taking reasonable steps to understand the nature of the operations and the hazards and risks associated with them (s 27(5)(b)). Where AI has materially changed how work is designed and how decisions are made, that change is now part of the operations an officer must understand.
Five levers for protecting cognitive health at work
MHI’s report sets out five levers for building brain capital at organisational and societal level: safeguard brain health, foster brain skills, study brain capital through better measurement, invest in it deliberately, and mobilise stakeholders around a shared agenda (Coe et al. 2026). Applied to cognitive health in an AI-enabled workplace, each has a practical reading.
Safeguard brain health.
Sleep, recovery and psychological detachment from work sit at the base of everything else here. A related McKinsey analysis is explicit that sleep and recovery underpin the brain’s capacity to consolidate learning and clear metabolic waste, and that cognitive capacity depletes with sustained effort (Brassey et al. 2026). An organisation layering AI-driven intensity onto an already depleted workforce is spending brain capital it has not budgeted for.
Foster brain skills.
The Baldeo finding points to a specific, teachable behaviour: active engagement with AI output, modifying it, challenging it, checking it, is associated with preserved confidence and ownership, while passive acceptance is associated with the erosion of both (Baldeo 2026). That is a behaviour that can be built deliberately, through training and through how work is designed, rather than left to individual habit.
Study and measure.
In consulting practice, it is common to meet organisations that can quote their AI adoption metrics in detail yet have no comparable visibility of the cognitive load or confidence trajectory of the people doing the work. The report’s measurement lever is framed at a societal level; extending it to organisational-level monitoring of cognitive demands is a natural application, and one a structured psychosocial risk assessment already points towards.
Invest.
The report’s economic case is that proactive investment in employee health, including brain health, can generate substantial returns, and it cites workplace intervention case studies with returns well above cost (Coe et al. 2026). For a duty holder, the calculation is simpler: the cost of protecting the judgement of the people doing safety-relevant and business-critical thinking is small against the cost of that judgement failing.
Mobilise.
MHI’s report argues brain capital needs to move from a human resources initiative to a chief executive and board-level priority (Coe et al. 2026). The WHS framing sharpens the point: where cognitive load is a psychosocial hazard, oversight of it falls within the officer’s due diligence obligation rather than sitting as optional strategy.
The discipline this requires
None of this is an argument against using AI. The evidence points to a narrower conclusion: active engagement with AI output holds up better than passive acceptance, and neither works without protected time to think.
Speed of adoption is not what will separate organisations here. What will separate them is whether they treated the cognitive health of their people as seriously as the technology, and recognised that the duty to manage it already exists.
Frequently asked questions
What is cognitive health at work?
Cognitive health at work refers to the state of the brain functions people rely on to do their jobs: attention, memory, reasoning, judgement, and the capacity to recover from mental effort. The McKinsey Health Institute and World Economic Forum describe it as one half of “brain capital”, alongside brain skills such as critical thinking and adaptability (Coe et al. 2026). In a work health and safety context, it covers how work design, workload, and technology affect a worker’s capacity to think clearly and make sound decisions.
Is cognitive load a hazard under Australian WHS law?
Yes. Under the model WHS Regulations, psychosocial hazards include hazards arising from the design or management of work, and duty holders must consider job demands and systems of work when determining controls (rr 55A, 55D). The model Code of Practice lists sustained concentration or vigilance, tasks that exceed a worker’s capacity, and rapid task-switching as examples of high job demands (Safe Work Australia 2022). Victoria imposes comparable duties under its own regulations.
Does using AI at work harm cognitive health?
The evidence does not support claims of brain damage. A peer-reviewed 2026 study found passive AI reliance is associated with reduced confidence in one’s own reasoning, while active engagement with AI output is associated with retaining it (Baldeo 2026). A separate preprint reported lower neural engagement among AI-assisted writers, but that study is small, not yet peer reviewed, and contested (Kosmyna et al. 2025; Stanković et al. 2026). The prudent reading is preliminary evidence worth managing, not a proven harm.
What should employers do when deploying AI?
Treat the deployment as a change to the design of work. That means consulting workers before implementation (model WHS Act s 49(d)), assessing the psychosocial risks created by changed job demands and new monitoring roles, protecting uninterrupted time for judgement-heavy tasks, and training people to actively check and challenge AI output rather than passively accept it.
Who is responsible for cognitive health at work?
The person conducting a business or undertaking holds the primary duty, including the duty to provide and maintain safe systems of work (model WHS Act s 19(3)(c)). Officers must exercise due diligence, which includes understanding how AI has changed the operations and their hazards (s 27(5)(b)). Workers retain duties to take reasonable care, but the design of cognitively safe work rests with the duty holder.
You might like to read WHS to benefit from AI
References
American Psychological Association 2026, Overreliance on AI programs may undermine confidence at work, press release, 16 April, viewed 2 August 2026, <https://www.apa.org/news/press/releases/2026/04/overreliance-ai-undermine-confidence>.
Bainbridge, L 1983, ‘Ironies of automation’, Automatica, vol. 19, no. 6, pp. 775-779.
Baldeo, S 2026, ‘Generative AI reliance and executive function attenuation: behavioral evidence of cognitive offload in high-use adults’, Technology, Mind, and Behavior, published online 16 April, doi:10.1037/tmb0000191.
Brassey, J, Goran, J, Krivkovich, A & Smaje, K 2026, ‘Five principles for designing brain-powered organizations’, People & Organization Blog, McKinsey & Company, 13 July, viewed 2 August 2026, <https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-organization-blog/five-principles-for-designing-brain-powered-organizations>.
Coe, E, Brassey, J, Enomoto, K & Pérez, L 2026, The human advantage: stronger brains in the age of AI, McKinsey Health Institute in collaboration with the World Economic Forum, 15 January, viewed 2 August 2026, <https://www.mckinsey.com/mhi/our-insights/the-human-advantage-stronger-brains-in-the-age-of-ai>.
Kosmyna, N, Hauptmann, E, Yuan, YT, Situ, J, Liao, X-H, Beresnitzky, AV, Braunstein, I & Maes, P 2025, Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing task, preprint, arXiv:2506.08872, MIT Media Lab, Cambridge, MA.
Mark, G 2023, Attention span: a groundbreaking way to restore balance, happiness and productivity, Hanover Square Press.
Mark, G, Gudith, D & Klocke, U 2008, ‘The cost of interrupted work: more speed and stress’, Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, Association for Computing Machinery, New York.
Occupational Health and Safety (Psychological Health) Regulations 2025 (Vic).
Parasuraman, R & Riley, V 1997, ‘Humans and automation: use, misuse, disuse, abuse’, Human Factors, vol. 39, no. 2, pp. 230-253.
Safe Work Australia 2022, Model code of practice: managing psychosocial hazards at work, Safe Work Australia, Canberra.
Safe Work Australia n.d., Model Work Health and Safety Act, Safe Work Australia, Canberra.
Safe Work Australia n.d., Model Work Health and Safety Regulations, Safe Work Australia, Canberra.
Stanković, M, Hirche, E, Kollatzsch, S & Doetsch, JN 2026, ‘Comment on: your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing tasks’, preprint, arXiv:2601.00856, University of Vienna and Technische Universität Dresden.
This article is general commentary on work health and safety matters and does not constitute legal advice. Duty holders should obtain advice on the application of the legislation in force in their jurisdiction to their specific circumstances.