I have been marking an assignment that asks students to use AI to generate a draft and then improve it through critical thinking. The irony is hard to miss. The case imagines a consulting firm planning to shift its graduate capability programme towards AI skills. Students are then asked to write a memo arguing against that move and to appeal instead to the firm’s ethics and values in defence of human competencies. The assignment is useful precisely because it captures the core tension of the present moment: AI can improve output, but that does not mean it is strengthening the human capabilities that matter most in professional life.
That is why the recent report by Lodge and Loble on artificial intelligence and cognitive offloading is so timely. Their key contribution is not a simplistic argument for or against AI. It is a more precise distinction between forms of offloading that support thinking and forms of offloading that replace it. In their account, the real risk is a performance paradox: AI may improve short-term task performance while weakening the deeper learning and schema formation that education is meant to develop (Lodge & Loble, 2026).
For me, this idea travels neatly beyond the classroom. In consulting, the biggest failure is often not poor presentation or weak formatting. It is solving the brief that was written rather than discovering the problem that actually needs to be solved. The written brief is rarely the whole brief. It is usually shaped by time pressure, politics, institutional habits, assumptions, and missing voices. AI can help process that artefact faster. It cannot, by itself, tell us what has been omitted, what is politically sensitive, whose interests are absent, or what success should really mean.
From cognitive offloading to the performance paradox
Cognitive offloading is not new. People have always used tools to reduce mental burden. Notes, calendars, calculators, and checklists all change what the mind has to hold and do. What is new is that generative AI allows users to offload not only memory or calculation, but also drafting, organising, comparing, summarising, and even the appearance of reasoning. Used well, this can be helpful. Used badly, it can lead people to outsource precisely the thinking they most need to develop (Lodge & Loble, 2026).
The most useful distinction is between beneficial and detrimental offloading. Beneficial offloading reduces lower-order or extraneous load. It might tidy prose, generate practice questions, or cluster interview themes so that the person can spend more time on interpretation. Detrimental offloading replaces intrinsic cognitive work such as analysis, synthesis, reflection, interpretation, and judgement. At that point, the output may still look polished, but the capability underneath begins to weaken (Lodge & Loble, 2026).
This is the performance paradox in practical form. Bastani and colleagues found that students using GPT-4 in high-school mathematics performed better during AI-assisted practice, but one version of the tool was associated with worse later performance when AI access was removed. The point is not that AI is always harmful. It is that poorly structured AI support can produce immediate gains while weakening independent performance later (Bastani et al., 2025).
Fluency on demand and the illusion of competence
One reason this matters is that AI generates fluent output very easily. It can produce prose that sounds articulate, structured, and confident. That fluency is often useful, but it also creates a misleading metacognitive cue. Users may begin to mistake the ease of generating a response for depth of understanding. What looks like competence can become a mirage of competence.
This is where the idea of metacognitive laziness becomes helpful. Fan and colleagues found that generative AI can promote dependency and potentially reduce self-regulated learning and deep engagement when learners rely on the tool rather than actively monitor, direct, and evaluate their own thinking. The danger is not simply that AI may be wrong. It is that people may become less inclined to notice when they themselves no longer understand enough to judge the quality of the output (Fan et al., 2025).
That distinction matters for assessment, but it matters just as much in professional work. A consultant who can produce polished wording with AI has not necessarily strengthened their judgement. A graduate who can submit a coherent memo with AI support has not necessarily learned how to think through a case. In both settings, apparent capability can outpace real capability.
The brief is not the whole brief
This is where the consulting angle becomes especially important. Most consulting failures do not happen because nobody can make a deck. They happen because the team never properly defined the problem in the first place. The brief may say one thing, while the real issue lies somewhere else: in stakeholder conflict, implementation fatigue, political sensitivities, power asymmetries, unclear success criteria, or symptoms being mistaken for causes.
Problem framing is therefore not simply an analytical step. It is a human process involving listening, ambiguity tolerance, context reading, relationship-building, and reframing. It requires consultants to notice what is not being said, who is not in the room, and which assumptions have quietly shaped the original request. AI can assist with preparation for this stage. It cannot responsibly replace it.
AI can accelerate preparation, but it cannot replace the human work of framing the right problem.
When efficiency becomes exhaustion
There is another reason to take this seriously. Poor AI use does not only risk weaker learning or weaker judgement. It may also make work more tiring.
Recent Harvard Business Review writing argues that AI often does not reduce work so much as intensify it. Tasks become faster, so more tasks are attempted. Expectations rise. Work expands into more hours of the day. The promise of efficiency becomes workload creep rather than relief (Ye, 2026).
A related Harvard Business Review article describes “AI brain fry” as mental fatigue produced by excessive use or oversight of AI tools beyond one’s cognitive capacity. Its key point is not the slogan but the organisational lesson: AI does not remove the limits of human attention simply because it increases the speed of machine output. Some patterns of AI use appear to increase cognitive fatigue, while better-designed workflows can reduce burnout instead (Bedard et al., 2026).
This helps explain a shift many professionals will recognise: work moves from making to reviewing. Creating from scratch can be demanding, but it often has coherence and flow. Reviewing streams of plausible but uneven machine output can be more fragmented, more vigilance-heavy, and more mentally draining. AI may save time on one task while consuming attention across many smaller supervisory decisions.
Siddhant Khare makes the same point from lived experience. His essay on AI fatigue is valuable not because it is formal research, but because it captures something many users are beginning to feel: productivity gains can coexist with greater exhaustion. Time saved does not automatically become time recovered. Sometimes it simply becomes more prompting, more checking, and fewer natural pauses.
A consulting table for thinking about AI use
This is the framework I find most useful. It is not a ban on AI. It is a way of deciding where AI belongs and where human capability must remain central.
| Consulting stage | AI should assist with… | Humans must remain responsible for… | Main capability risk | Main fatigue risk |
|---|---|---|---|---|
| Problem framing | Background scanning, summarising prior material, surfacing assumptions, first-pass issue trees | Defining the real problem, reading stakeholder tensions, clarifying purpose, judging what matters | Solving the wrong problem because the written brief is mistaken for the whole brief | False confidence from quick synthesis |
| Research | Search-term generation, document triage, thematic extraction, first-pass summaries | Source evaluation, credibility judgement, synthesis, deciding relevance | Shallow evidence standards and over-trust in summaries | Review overload from checking many partial outputs |
| Analysis | Data sorting, clustering, pattern spotting, scenario generation, option generation | Interpretation, testing causality, weighing trade-offs, contextual judgement | Outsourced reasoning and false pattern recognition | Reviewer fatigue from plausible but weak outputs |
| Client communication | First-pass email drafts, note organisation, structure suggestions, alternative phrasings | Tone, empathy, relationship judgement, political sensitivity, trust-building | Declining independent communication capability | Constant editing and context switching |
| Recommendations | Scenario modelling, option generation, draft structures, pros-and-cons lists | Prioritisation, feasibility judgement, accountability, ethical reasoning | Polished but context-poor advice | Mental overload from comparing endless generated options |
| Risk mitigation | First-pass risk registers, red-flag prompts, stress-testing scenarios | Judging severity, escalation, governance, human and ethical review | Missing cultural, political, or human-development risks | Vigilance fatigue |
| Implementation planning | Draft timelines, task breakdowns, sequencing options, communication-plan scaffolds | Realism, readiness assessment, adaptation, change leadership | One-size-fits-all rollout logic | Planning sprawl and hidden coordination load |
| Learning and capability building | Reflection prompts, critique prompts, practice-question generation, rehearsal support | Reflection, self-explanation, judgement formation, professional growth | Long-term dependency and weaker independent capability | Cognitive passivity disguised as efficiency |
The distinction underneath this table is straightforward. AI is most useful when it offloads preparatory, lower-order, or exploratory work. It becomes risky when it offloads problem framing, interpretation, judgement, ethical reasoning, and relationship work. That is not only a pedagogical distinction. It is a professional one.
From framework to practice: the AI Use Simulator
To make this more practical, I have also built an AI Use Simulator. The point of the tool is not to ask whether AI is being used, but whether it is being used in a way that supports thinking or replaces it. It presents scenarios from education, consulting, and professional life, and asks a more useful question: what exactly is being offloaded, and with what capability effect?
That matters because not all AI use is equivalent. A student using AI to improve grammar is doing something very different from a student using AI to perform the analysis they were meant to learn. A consultant using AI to summarise a brief before stakeholder interviews is doing something very different from a consultant accepting that summary as the problem itself. The simulator is not an anti-AI device. It is a judgement device.
Why this strengthens the case for human competencies
This is also why the discussion links so strongly to the human competencies that matter in an AI-rich world. In my earlier work on the 12 critical skills for the 21st century, I argued that capabilities such as communication, cultural humility, dealing with ambiguity, critical thinking, emotional intelligence, creative thinking, ethics and sustainability, AI literacy, and adaptability are essential for navigating a complex and rapidly changing world. The argument feels even stronger now. These are not residual skills left behind after AI has done the “real work”. They are what determine whether the work was framed, interpreted, communicated, and governed properly in the first place.
Critical thinking tests the brief. Creative thinking reframes it. Communication and emotional intelligence help people hear what is missing. Cultural humility helps them recognise the limits of their own frame. Dealing with ambiguity helps them resist premature closure. AI literacy helps them decide what can be delegated safely and what should not be. Ethics and sustainability help ensure that speed is not purchased at the cost of human capability or human wellbeing. These are exactly the kinds of capacities firms should be strengthening in graduates, not trading away in pursuit of a narrow reading of technological relevance.
There is also a second tool worth linking here: the competencies reflection tool. There is also a second tool worth linking here: the competencies reflection tool. At its best, education is not only about performance. It is also about self-discovery — learning what kind of thinker, collaborator, and decision-maker one is becoming. Knowing one’s strengths matters, but so does recognising one’s blind spots, default habits, and unfinished capacities. That is not just a graduate attribute. It is part of lifelong learning. In an AI world, where difficulty can be deferred with a prompt, reflective awareness of one’s own capability profile becomes even more important. A tool that supports that kind of self-discovery belongs at the heart of this conversation.
The real professional question
The real question is not whether AI can help people work faster. Of course it can. The more important question is whether our use of AI is building the capacities that matter when conditions are ambiguous, stakes are real, and the written brief is incomplete. If AI use leaves students less able to reason independently, consultants less able to frame problems, and professionals more fatigued from constant review and switching, then the apparent productivity gain is a poor bargain.
Used well, AI can extend capability. Used badly, it can intensify work while hollowing out the very capacities professionals need most. That is why the goal should not be less thinking. It should be better allocation of thinking. And that is why the human bottleneck is not a defect to be engineered away. It is often the place where judgement, responsibility, and real problem definition still occur.
References
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Bedard, J., Kropp, M., Hsu, M., Karaman, O. T., Hawes, J., & Rosen Kellerman, G. (2026, March 5). When using AI leads to “brain fry”. Harvard Business Review. https://hbr.org/2026/03/when-using-ai-leads-to-brain-fry
Fan, Y., Tang, L., Seeshing Yeung, S., Tong, C., Meng, Y., Dong, Y., Cui, Y., Kuang, Y., & Gasevic, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes and performance. British Journal of Educational Technology. Advance online publication. https://doi.org/10.1111/bjet.13544
Khare, S. (2026, February 7). AI fatigue is real and nobody talks about it. Siddhant Khare. https://siddhantkhare.com/writing/ai-fatigue-is-real
Lodge, J. M., & Loble, L. (2026, March). Artificial intelligence, cognitive offloading and implications for education. University of Technology Sydney / Network for Quality Digital Education. https://doi.org/10.71741/4pyxmbnjaq.31302475
Ye, X. M. (2026, February 9). AI doesn’t reduce work—It intensifies it. Harvard Business Review. https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies-it