He’s more machine now than man*
AI may make us perform better before it makes us better.
*Obi-Wan Kenobi
AI is extraordinarily useful! Let’s just put that on the table from the outset. And it has nothing to do with the ‘force’ or the ‘dark side’! It helps us draft, summarize, analyze, translate, compare, classify, code, research, ideate, polish etc. etc. It allows people to move faster and produce work that is often clearer, sharper and more complete than anything they could have produced unaided.
When the output improves, it is tempting to assume the person has improved. A better deck created by a better consultant. A sharper argument delivered by a sharper thinker. A clearer strategy document implies a clearer strategic mind.
A recent paper by Emma Wiles and colleagues[i] suggests that this assumption may be wrong. It describes generative AI as an “exoskeleton” for knowledge workers. It is a useful metaphor because it captures both the promise and the danger. An exoskeleton can help someone lift more than they could alone, extending strength, compensating for weakness, and allowing people to perform tasks beyond their unaided capability. That is not trivial benefit. In many contexts, it could be transformative.
But wearing an exoskeleton is not the same as building muscle. AI may improve immediate performance without necessarily developing durable human capability. It may help people produce stronger artefacts without strengthening the underlying habits of thought, judgement and practice that those artefacts appear to represent. In other words, the danger is not simply that AI makes us lazy, but rather it makes us look capable before we have become capable.
This matters acutely in professional practice because so much of what we call competence is inferred from outputs. We judge people by the quality of their slides, reports, emails, analyses, recommendations and presentations. They are the visible manifestations of the work. But they are not the whole of work. A good consultant is not just someone who can produce a convincing strategy document. A good lawyer is not just someone who can draft a plausible argument. A good analyst is not just someone who can summarise a market. A good leader is not just someone who can express a clear point of view. In each case, the artefact matters, but so does the process behind it: the framing of the problem, the interpretation of context, the weighing of evidence, the sensing of politics, the recognition of uncertainty, the ethical judgement, the ability to defend a position, and the willingness to change one’s mind when the facts demand it.
Generative AI is increasingly good at helping with the artefact. It is less clear that it automatically develops the process or our innate ability as practitioners.
This is where Frank Blackler’s[ii] old distinction between knowledge and knowing becomes useful. Knowledge is often treated as a thing: something stored, transferred, retrieved, documented or possessed. But knowledge work is not just the manipulation of informational objects. It is an active process of knowing. It is situated, social, interpretive and directed towards action. We know through practice, through language, through collaboration, through routines, through challenge, through experience, and through the messy interaction between what is written down and what is lived.
AI is very powerful at the encoded layer of work: words, documents, data, models, summaries, classifications, arguments. It can take a vast amount of encoded material and produce a synthetic output with impressive speed. That is part of its power. But professional capability also depends on forms of knowing that are harder to encode … what Blackler calls embrained knowledge (conceptual skill and judgement); embodied knowledge (rooted in experience and contextual clues), encultured knowledge (rooted in shared language and interpretation) and embedded knowledge (held in routines and systems)
The exoskeleton works best where knowledge has already been made explicit. It is less reliable where competence depends on reading the room, sensing what matters, spotting what is missing, understanding how a client will hear something, or knowing when a technically correct answer is practically useless. This creates a new organisational risk: the illusion of capability.
The capability illusion appears when better AI-assisted outputs are mistaken for stronger human capability. The organisation sees improved production and assumes improved development. It sees greater fluency and assumes greater understanding and judgement.
The risk is especially pronounced for junior knowledge workers. For decades, junior people have learned partly through friction. They have had to wrestle with messy source material, produce weak first drafts, build clumsy hypotheses, get challenged by seniors, revise their thinking, and slowly develop the ability to distinguish the important from the merely available. Much of that process is inefficient. It is also developmental.
AI can smooth that path. It can provide the first structure, the first summary, the first draft, the first set of implications. Again, this is useful. No one should romanticise needless drudgery. But we should be careful about removing all the struggle from the early stages of expertise. The struggle is often where the learning happens.
A junior consultant who asks AI to summarise twenty interviews may save hours. But if they never sit with the raw mess of what customers actually said, they may miss the hesitation, contradiction, emotion and oddness that make insight possible. A manager who asks AI to draft a strategic recommendation may get a better document, but if they never form their own provisional view before reading the machine’s answer, they may not develop the internal discipline of judgement. A student who uses AI to produce a fluent essay may submit stronger work, but that very fluency may conceal the thinness of their understanding.
This is not an argument for banning AI from the learning process. That would be both futile and foolish. The question is not whether people should use the exoskeleton. The question is whether using the exoskeleton helps them become stronger, or merely allows them to perform as though they were strong.
Fernandes and colleagues[iii] also make this point in their work on AI, performance and metacognition. Their argument, captured in their blunt title “AI makes you smarter, but none the wiser,” is that AI can improve task performance while leaving people poorly calibrated about their own ability. That is exactly the danger in knowledge work. People may produce better outputs without becoming better judges of their own thinking. Worse, they may become more confident because the work looks better.
This is where the capability illusion becomes self-reinforcing. Better AI-assisted outputs generate praise. Praise generates confidence. Confidence reduces scrutiny. Reduced scrutiny increases dependence. Dependence becomes normal. Before long, the organisation is not simply using AI to extend capability …. It is assuming a capability it may not actually possess.
There is a leadership problem here. Many organisations are measuring AI adoption through visible productivity gains: more, faster, cheaper. These metrics are not unimportant. But they are incomplete. They tell us what the human-AI system can produce. They do not tell us what the human is learning.
This distinction matters because systems can fail in unfamiliar conditions. The exoskeleton may work brilliantly on routine or semi-routine tasks. It may be excellent when the question is clear, the evidence is available, the domain is familiar, and the cost of error is low. But in ambiguous, novel or high-stakes situations, beyond what Dell’Acqua et al.[iv] would call the “jagged technological frontier”, the human still needs to understand what is going on. They need to know what the machine has missed, where it has over-smoothed reality, and when its neat, fluent answer has become a trap.
This is why AI capability programs that focus only on prompting are inadequate. Prompting is useful, but it is not the core human capability. The deeper capability is knowing how to think with the tool without letting the tool think for you. It is knowing when to ask for help, when to challenge the answer, when to go back to the source material, when to seek another human view, and when to reject the elegant synthesis because it has flattened the problem.
That means designing AI use so that it builds human capability rather than simply supplying machine-assisted performance. This is where recent work by Drosos and colleagues[v] is interesting. Their research on AI “provocations” explores how systems can be designed to make users think: by challenging assumptions, offering counter-arguments, surfacing alternatives, or prompting reflection rather than simply completing the task. The point is important. AI does not have to remove friction. It can create productive friction.
This is the design challenge for organizations. Most current AI deployment is built around convenience. The user asks; the system answers. The interface is optimised for smoothness, but learning often requires resistance. It requires being asked: What do you think before you ask the machine? Why do you believe that? What evidence would change your mind? What has been omitted? Where might this answer fail? What would a sceptical client say? What would someone closer to the customer notice? What are you prepared to stand behind? In this model, AI becomes more than an answer engine. It becomes a thinking partner, a critic, a simulator, a coach and a mirror. It does not merely accelerate output, it strengthens the human processes behind the output.
For leaders, this means asking a different set of questions about AI adoption. Not just: are people using it? Not just: are they producing more and/or are we saving time? The more important questions are: are people becoming better at framing problems? Are they becoming more discriminating? Are they better able to judge quality? Are they more aware of uncertainty? Are they learning faster? Are juniors developing the underlying craft, or simply producing more senior-looking outputs? Are teams preserving the forms of knowing that do not live neatly inside documents and datasets?
There is a useful test. Remove the exoskeleton, even briefly. Ask people to explain the reasoning behind the work. Ask them what they would do if the AI answer were unavailable. Ask them to defend the recommendation under challenge. Ask them to identify the assumptions. Ask them to describe what they learned from the process, not just what they produced at the end of it.
The goal is not nostalgic human purity. Knowledge work has always involved tools … we think with spreadsheets, search engines, colleagues, templates, models and methods. Generative AI is another extension of that long human habit of building cognitive scaffolds around ourselves. But is the scaffold helping us build or merely holding us up. We are reminded of Scottish writer, Andrew Lang’s great analogy that one often: “uses statistics as a drunken man uses lamp-posts—for support rather than illumination.”
That is the difference organisations now need to notice. If AI is used only to produce more polished artefacts, it may create a generation of knowledge workers who look increasingly capable while becoming increasingly dependent on systems they do not fully understand. If AI is used well, however, it can become a powerful developmental environment: one that exposes people to new possibilities, challenges their assumptions, stretches their thinking and helps them practise judgement.
The exoskeleton is not the enemy. The illusion of capability is the issue. The real prize is not work that looks smarter. It is people who become wiser in the process of doing it.
The idea for this post came from reading a recent FT article ‘Is AI an exoskeleton for the mind?’ (June 30, 2026) which cited the Wiles et al research
[i] Wiles, E., Krayer, L., Abbadi, M., Awasthi, U., Kennedy, R., Mishkin, P., Sack, D., and Candelon, F. 2024. GenAI as an Exoskeleton: Experimental Evidence on Knowledge Workers Using GenAI on New Skills. Available at SSRN: https://ssrn.com/abstract=4944588 or http://dx.doi.org/10.2139/ssrn.4944588
[ii] Blackler, F. 1995. Knowledge, knowledge work and organizations: An overview and interpretation. Organization Studies, 16(6), 1021–1046.
[iii] Fernandes, D., Villa, S., Nicholls, S., Haavisto, O., Buschek, D., Schmidt, A., Kosch, T., Shen, C. and Welsch, R., 2026. AI makes you smarter, but none the wiser: the disconnect between performance and metacognition. Computers in Human Behavior. Volume 175
[iv] Dell’Acqua, F., McFowland, III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Technology & Operations Mgt. Unit Working Paper No. 24-013.
[v] Drosos, I., Sarkar, A., Xu, X., & Toronto, N. (2025). “It makes you think”: Provocations help restore critical thinking to AI-assisted knowledge work. arXiv preprint arXiv:2501.17247.