The shortcut paradox: how AI is changing the way expertise is built

Artificial intelligence, expertise, machine learning

AI saves us time. But by taking over part of what we used to do ourselves, it is also changing the way we learn. And expertise is forged precisely through trial, error and accumulated experience. So how do we train tomorrow’s skilled professionals in an age of shortcuts?

Founded in 2025 and valued at $2.3 billion, General Intuition trains its AI on hundreds of millions of hours of video games, analysing what players see and the decisions they make. The aim: to teach machines to find their bearings, anticipate and act in unfamiliar environments. There’s a curious irony here: to build expert machines, we start by feeding them experience. Meanwhile, at work, we use AI precisely to skip some of that trial and error.

Yet it is exactly this accumulation that eventually produces the famous professional ‘eye’: the one that allows an oenologist, a mountain guide or a perfumer to spot almost instantly that a situation ‘doesn’t feel right’. They aren’t guessing: they’re recognising. Which raises a question: how do we train the experienced professionals of tomorrow if AI lets us skip part of the journey towards that professional intuition?

When experience becomes intuition

In the professional world, intuition often seems less legitimate than data or a method. Yet expert intuition is built: through repeated experience, certain signals become almost instantly recognisable. Thierry Denjean, an entrepreneur who owns several wine estates, learnt this without any training as an oenologist. Some thirty years ago, he learnt the ropes alongside a winemaker friend in Bandol, in Provence, by observing the soil, the vines and the climate. It was learning by immersion far more than by theory. The same goes for Blaise Agresti, a high-mountain guide and former head of mountain rescue units. The son of a guide, he grew up immersed in that environment: ‘I was exposed to the natural elements from a very young age,’ he recalls. Climbing, nights outdoors, sheer drops: even before his professional training, his eye was already being shaped by experience.

  • Experts don’t guess, they recognise

This is exactly what the researcher Herbert Simon describes: expert intuition rests on recognising situations we have already encountered. Behind its apparent spontaneity lies a vast library of experience. The work of psychologist Gary Klein with fire commanders points in the same direction: the most experienced often recognise a familiar pattern and a plausible response, without weighing up several scenarios.

For Thierry Denjean, this library of experience was built through repetition. During the harvest, the team tastes the vats, tries out blends, adjusts the proportions, then starts again. Over time, certain differences start to jump out. For Blaise Agresti, expertise was also built step by step. During his first major operations, he says he was ‘reacting rather than relying on intuition or anticipation’. Then, as one expedition followed another, ‘you start to get a better feel for things’. Gradually, certain signals become easier to read: what looks like a hunch is often the product of accumulated experience that has become hard to put into words.

  • A seasoned eye

With experience, what was invisible at first eventually becomes a signal. This is one of the lessons of Hubert and Stuart Dreyfus’s model of skill acquisition : novices start by relying on explicit rules, then gradually learn to prioritise information and recognise what matters. The time it takes to become a high-mountain guide illustrates this process. Blaise Agresti mentions four to five years of practice before the entrance exam, then as many years of training. ‘In fact, it’s a cycle that takes a good ten years,’ he sums up. A long time to build up situations, decisions, mistakes and their consequences.

In the vineyard, Thierry Denjean compares this ability to a doctor’s diagnosis: ‘If they don’t have that ability to observe… It’s the same with vines.’ Reading a leaf, getting a feel for the soil, spotting that a plant is struggling: it is the repetition of these acts, and seeing them in context, that trains the eye. This is also where AI shows its limits. You can show it a photo of a leaf, but it misses part of the context: how deep the roots go, exposure to the wind, the type of soil, recent weather. ‘All of that takes a lot of observation.’ But if AI takes over some of the very tasks that forced us to observe, compare and interpret, how will novices, in turn, learn to see?

Satnav syndrome: the cognitive cost of shortcuts

Satnav is a good illustration of this mechanism: it guides us efficiently, but partly spares us from learning the way. A study published in 2020 in Scientific Reports links heavier GPS use with poorer spatial memory.

  • When AI short-circuits the learning loop

Psychologists call it cognitive offloading : relying on an external aid to reduce the mental effort a task requires. Generative AI takes this further by handling part of the research, summarising, reasoning or writing.

Yet learning depends precisely on not getting the answer straight away. Nawal Abboub, who holds a PhD in cognitive science and teaches about the uses of AI at ENS-PSL, points out that learning goes through several stages: selecting information, testing a hypothesis, getting it wrong, receiving feedback, then trying again. ‘That’s what allows us to build a representation of that knowledge,’ she explains.

With AI, the path can be shortened: ask → get → check → deliver. The time saved is obvious, but ‘you haven’t gone through all those stages of acquisition’, Nawal Abboub stresses. And our brains are only too happy to oblige: ‘Why remember a phone number when you have a database that holds everything?’ she says.

  • Producing is not learning

This is the crux of the problem: getting a good answer with AI does not mean you have acquired the skills to produce it or check it. Quite the opposite: ‘The more sophisticated our tools, the more sophisticated our skills need to be,’ Nawal Abboub insists. The better the tool gets, the more users need to know how to question it, put it in context and spot its mistakes.

The problem is particularly acute for beginners: AI can help them quickly produce a result close to that of an experienced professional, before they have built the knowledge needed to assess it. Yet, as she points out, ‘you can’t think critically without knowledge’. In an article published in 2026 in Trends in Cognitive Sciences, Trent Cash and Megan Kelly also note that delegating certain tasks to AI can slow down skill acquisition or speed up skill erosion, depending on how the tool is used.

We are probably facing an unprecedented learning equation: how do we develop solid experts if part of the journey is now taken over by the machine ?

At Le Connecteur, we cultivate what Lia cannot replace

How can we reinvent the way experience is built?

If AI can short-circuit part of the learning journey, one answer is to rethink the conditions in which we learn with it.

  • From answer machine to AI tutor

For Nawal Abboub, it all comes down to the stance we adopt. ‘There’s a passive way and an active way,’ she explains. ‘Instead of asking it for the result, ask it for the route.’ Used as a smart tutor, AI can ask questions, flag a mistake, offer a hint, suggest an intermediate step or prompt you to rephrase your reasoning. Medicine is already testing this approach with virtual patients powered by large language models. A study published in 2025 shows that they can help train clinical reasoning by exposing students to a much wider range of situations.

  • Use your own neurons… before the AI’s

At Epitech, a school of computer science, this principle lies at the heart of its teaching. Pierre Robert, a learning designer and author of Déchaîner l’intelligence (FYP éditions), advocates project-based learning: students search, test and make mistakes, while teaching staff ask questions without handing over the solution. His motto: ‘Every bit of help you give is a piece of learning stolen.’ Yet the arrival of LLMs did not lead the school to ban them. ‘We allow AI, and we even encourage students to use it,’ Pierre Robert explains. But with one rule: ‘Start with your own brain before you use the AI’s neurons.’ Epitech sums up this approach in three steps: raise awareness, build responsibility, make it last. In other words: understand cognitive debt, preserve reasoning, then check that the skill holds up over time.

  • No longer mistaking a deliverable for a lasting skill

It is the third principle, ‘make it last’, that brings about the deepest change. Until now, producing a working deliverable was largely enough to show you knew how to do something. But that link can no longer be taken for granted. ‘You can produce things very easily without any skills at all,’ Pierre Robert observes. A student can now generate a website or hundreds of lines of code without understanding what is going on behind the scenes.

So Epitech now spreads part of its assessment over time: several months after a project, students may be questioned again to check what they have retained. ‘Skill acquisition is no longer about proof of a deliverable. It’s about retention over the medium term.’ The aim: to measure not just what has been produced, but what has actually been learnt.

  • Starting from the job to rebuild learning

For companies, the challenge is to start from the job itself: which tasks should be delegated to AI, which should still be practised, and what knowledge must be preserved to judge its answers? Nawal Abboub notes that some teams are already coming up with their own safeguards: ‘You also need to listen to people on the ground, because they find solutions.’ Epitech applies the same thinking to its teaching support: AI flags certain unusual choices in students’ code, so that teaching staff can focus on the points worth questioning and check that they are properly understood.

What if AI could also speed up the experience curve ?

In their study Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,179 customer support agents. Productivity rose by 14% on average, but by 34% among novice and lower-performing agents. Their hypothesis: AI spreads some of the best agents’ practices more quickly.

AI can short-circuit learning when it hands over the answer directly, or speed it up when it helps people practise and improve. The challenge, then, is not to refuse shortcuts, but to preserve the experiences that are essential to building expertise. In the age of AI, the question remains: which detours are still worth taking ?

signature block laure girardot

Share this news on :