I’ve been writing about AI and expertise for a while now, circling around a few central ideas about how we teach for, develop, and communicate expertise. In Three Dimensions of Expertise for AI, I argued that useful work with GenAI requires some combination of domain, technological and situated expertise. In Expertise Not Included, I made the related point that the people most likely to benefit from AI are often those who already know enough to judge its outputs. More recently, in Where does expertise live?, I tried to move beyond the idea of expertise as something contained entirely inside an individual’s head. Expertise is also social, contextual and conferred: it involves not only what you know, but where, when and on whose behalf you’re entitled to act.
Until recently, if you wanted expert-level work, you had two broad options: spend years developing the relevant expertise, or find someone who already had it. So we hired lawyers, consulted doctors, employed accountants, read books, completed courses and worked alongside more experienced colleagues. Expertise was scarce because it took time; however, GenAI has introduced a third option: rent it.
You can now ask a general-purpose AI system to draft legal language, analyse a financial report, explain a scientific concept, write computer code or help design a lesson. The quality varies, sometimes dramatically, but the output can be remarkably fluent, and it frequently carries the surface signals of expertise: disciplinary vocabulary, orderly reasoning, confident explanation and appropriately formatted conclusions. You can signal expertise without having accumulated much of the underlying capability yourself. But access to expertise is not the same as possessing it.
Does AI make us lazy?
It’s tempting to go looking for a single definitive study proving that GenAI either improves learning or destroys it. Of course, this technology (in the sense of consumer-facing LLM-based chatbots and related applications) is barely four years old and the evidence doesn’t support certainty in either direction. Research in this area is new, frequently based on out-of-date models, and methodologically uneven. Comparable studies use different platforms, different prompts, different subjects and different definitions of learning. Some measure the quality of an artefact (essay, exam, report, short-answer) produced while AI is available. Others assess knowledge immediately afterwards. Some examine what hangs around weeks later, whether knowledge transfers to some other context, or what happens when access to the AI is removed.
A useful starting point is the distinction made by Lixiang Yan, Samuel Greiff, Jason Lodge and Dragan Gašević in their 2025 Nature Reviews Psychology comment, “Distinguishing performance gains from learning when using generative AI”. GenAI can improve what a learner produces during an activity, they argue, without necessarily supporting the deeper cognitive and metacognitive activity associated with learning. Their point is that performance and learning shouldn’t be treated as interchangeable.
In the relatively large experiment documented in Generative AI without guardrails can harm learning, researchers gave nearly 1,000 high-school mathematics students in Türkiye access to one of two GPT-4 systems, built using the OpenAI API. One resembled an unrestricted ChatGPT interface. The other had teacher-designed guardrails that encouraged hints and guided assistance rather than simply providing answers. Both systems substantially improved students’ performance during practice. But when the AI was removed, students who had used the unrestricted version performed 17 per cent worse than students who had never had access to it at all. The negative effect was largely mitigated in the guarded tutor condition, although the guarded system didn’t produce a clear positive learning effect once the AI was removed either.
A smaller study by André Barcaui, ChatGPT as a cognitive crutch, followed undergrad students learning about AI. On a surprise test 45 days later, the students assigned to unrestricted ChatGPT use scored an average of 57.5 per cent, compared with 68.5 per cent among those assigned to traditional study methods. Both studies essentially raise the same question: what’s left after the immediate appearance of competence has faded?
Other experiments turn up variations of the same tension. In Beware of metacognitive laziness, 117 university students completed a writing task with support from ChatGPT, a human expert, a writing analytics tool or no additional tool. The ChatGPT group improved its essay scores, but didn’t show significantly greater knowledge gain or transfer. And in Cognitive ease at a cost, 91 university students used either ChatGPT 3.5 or Google to investigate nanoparticles in sunscreen. The ChatGPT group reported lower “cognitive load”, but produced weaker reasoning and argumentation.
These findings converge around three related but distinct ideas:
The first strand is the gap between performance and learning. What someone can produce with assistance is not necessarily what they can subsequently produce, explain or transfer independently.
The second is cognitive offloading. Risko and Gilbert define cognitive offloading as using physical action or features of the environment to reduce the mental demands of a task. The important question is not whether we offload cognition (we always have) but what we offload, under what conditions, and whether the capability we need is still being developed or maintained.
The third strand is an illusion of competence. Long before GenAI, Matthew Fisher, Mariel Goddu and Frank Keil argued that searching online for explanations could inflate people’s estimates of their own internal knowledge. Access to information became difficult to distinguish from possession of it.
GenAI – specifically in this context using LLM-based applications and chatbots to access information – intensifies that ambiguity. Search engines usually return documents, fragments and competing sources. A chatbot returns an integrated, personalised, conversational answer, and explains itself in the same window in which we formulated the question. The distance between asking and knowing appears to collapse, and it feels less like retrieving someone else’s knowledge and more like thinking.
What is expertise, anyway?
If expertise were simply the possession of information, AI would pose relatively little philosophical difficulty. Information is cheap, and has been getting cheaper over time. But expertise has rarely been defined as information alone.
Hubert and Stuart Dreyfus’ model, developed in works including Mind Over Machine, describes movement from the rule-bound behaviour of the novice towards the context-sensitive judgement of the expert. The expert doesn’t simply possess “more rules”. Through experience, they become better able to perceive what is important in a given situation. Rules don’t disappear entirely, but they start to recede behind a more holistic and responsive form of action. I used Dreyfus and Dreyfus’ model in applying the three dimensions of expertise to professional learning, arguing that we don’t need to become experts in everything, but we should have staff within an organisation who are always moving towards the centre and deepening their contextual (situated) expertise.
In a similar fashion, Michael Polanyi suggested that “we can know more than we can tell”. Expert knowledge includes traditions, inherited practices, implied values and forms of judgement that can’t be fully rendered as explicit instructions. An LLM can help a user imitate the language of expertise, but without either the user or the model having undergone that socialisation.
This brings me back to the distinction between techne and metis that I explored in Expert Signals. Techne is formal, codifiable and transferable. It can be expressed in procedures, manuals, textbooks and training datasets. Metis is practical, embodied and local. It’s the teacher reading the room, the gardener feeling the soil, or the tradesperson noticing that something about a supposedly routine job isn’t quite right.
Large language models are engines of techne. They generate patterns from enormous collections of data. They can also describe metis, and imitate the language of someone who possesses it. What they can’t provide is the user’s own accumulated, embodied experience of acting within a particular context.
Expertise, from this perspective, includes more than producing an acceptable answer. It involves noticing, discriminating, interpreting, improvising and deciding. It includes knowing when a familiar rule no longer applies. It includes membership of a community, recognition from others and accountability for what happens next. A model can generate a plausible clinical recommendation, but the clinician remains responsible for the patient. An AI model can produce a lesson plan, but the teacher still has to walk into the room.
Correct output is part of expertise, but so is the capacity to act when the output is incomplete, the situation changes, the available options are all imperfect, or the consequences belong to you.
Rented expertise
This is where I think rented expertise becomes useful as a concept:
Rented expertise is access to the products and performance of expertise without equivalent accumulation of the underlying knowledge, judgement, situated experience or responsibility.
Like any rental, it provides utility without ownership. That isn’t automatically a bad thing, since renting is often more efficient or expedient than owning. In terms of expertise, I don’t need to qualify as an accountant to lodge a tax return, train as a mechanic to maintain a bike, or become an expert in every discipline that touches my day-to-day life. We’ve always distributed knowledge, relied on specialists and paid for access to capabilities that individuals don’t possess. In one sense, we have always “rented” expertise.
The difference is that when I consult a human expert, the location of the expertise is usually visible. When I call a plumber (not my dad, he’s retired), I know that the plumber possesses a capability that I don’t. There’s a relationship between us, however brief. And in that brief interaction I am receiving the products of someone else’s expertise, and I know it.
GenAI blurs that boundary. These platforms are deliberately designed to give the user the impression of collaboration, but they often go even further, offering to “rewrite this in your style” or “make sure that the ideas are in your unique voice”. I don’t think any current LLM is actually capable of doing that – certainly, every time I experiment with AI writing in “my voice” I read it back as painful Claudespeak, and immediately delete most of it. Even so, it feels more like those words are originating from a close, adjacent space, rather than a totally different person. Many people who use AI a lot report that it is increasingly difficult to tell where your ideas stop and the machine’s begin.
This is why rented expertise sits at the intersection of the three strands from the research: performance, offloading and illusion. It is possible to simultaneously, and even unthinkingly, perform the expertise and learning required, offloading vital parts of the process and producing something which, to an outsider, gives the illusion of competence.
The line between renting and acquiring knowledge depends partly on participation. Am I being asked to make a judgement, or just accept one? Do I have to retrieve and apply knowledge, or only recognise a plausible response? Can I explain why the answer is appropriate? Can I transfer the reasoning to a different problem? Can I detect an answer that’s fluent but subtly wrong? Can I still act when the model is removed? And, most importantly: am I willing and able to accept responsibility for the result? These questions tell us far more than whether someone used AI at all.
Want to learn more about GenAI professional development and advisory services, or just have questions or comments? Get in touch:
Leave a Reply