AI Deskilling: What We Lose When We Stop Practising
In Brief
- Researchers use the word deskilling for three separate things: experts losing skills they stop using, trainees never building them in the first place, and work being simplified on purpose so cheaper staff can do it.
- A leading children's brain surgeon argues strongly for AI in medicine, but for narrow, specific questions. Assembling the whole operation stays a human job.
- The skills worth protecting are the ones you need on the day the tool fails or meets a case it was never built for.
- In high-stakes work, trust runs between people. A surgeon cannot tell a family that the AI had an off day.
- AI products are built for stickiness. Dependency is the business model, not a side effect.
AI deskilling is the loss of ability that follows when a capable tool takes over work you used to do yourself. The BBC News programme AI Decoded put that problem to three people who see it from different angles: Owase Jeelani, a consultant paediatric neurosurgeon at Great Ormond Street Hospital in London; Tapani Rinta-Kahila, a researcher at the Hanken School of Economics in Helsinki who studies how digital technology changes human work, expertise, and decision-making; and Stephanie Hare, author of Technology Is Not Neutral. Their answer was not a clean yes or no. Skill fades when the tool takes over the practice, and the tool earns its place when the question is narrow and the human still owns the plan around it.
The episode's framing term is cognitive offloading, the habit of handing thinking over to a machine instead of doing it in your own head. Presenter Marc Cieslak, the BBC's AI correspondent, opened with what he sees week after week. Tech companies claim that some new task can now be finished with a carefully worded prompt, work that used to take expertise and years of experience. The disagreement on the panel was never about whether that is happening. It was about which tasks you can safely hand over, and what you owe the work in return.
What researchers mean by AI-driven deskilling
Cieslak put it to Rinta-Kahila that the word deskilling sounds dramatic, and asked what researchers actually mean by it. His answer was that the term carries several different meanings, and he set out three of them.
The first is the one most people picture. You are already an expert, you lean too much on automation, and over time you lose skills you once had. Rinta-Kahila calls this skill erosion, and he treats it as an unintended side effect of the technology rather than anyone's plan.
The second is quieter and harder to reverse. You are not an expert yet, you are training to be one, and the training now runs through sophisticated automated tools. In his words, you "never really grapple the real kind of nitty-gritty part of the work", so you never really become that skilled. Nothing eroded here, and nobody planned it. The skill was simply never laid down.
The third is the one somebody chooses. In work-organisation research, deskilling means deliberately simplifying a task, using automation so that less skilled workers can do it. Rinta-Kahila spelled out the point of the exercise: you can then hire less skilled workers, who are typically cheaper labour.
Why paediatric neurosurgeon Owase Jeelani argues for AI assistance
Jeelani's opening position was blunt: AI assistance is "hugely, hugely important". He knew it was a strong stance in a programme about erosion, so he explained the reasoning behind it.
The public conversation, he argued, spends most of its attention on the harm AI can do to humans. He does not deny that potential. What he thinks gets too little attention is the harm the human mind already does to humans. Used properly, at the right level and on the right questions, AI can improve human thinking enormously, in his view. The trick is getting the balance right without leaning on the tool too much.
To explain the scale of the change, he turned to a much older technology. When humans discovered fire, a great deal changed, not only in how people lived but in their physiology, and humans adapted to the technology they had made. He expects AI to do something similar. The question he cares about is not whether the change happens but who steers it: how we monitor that change, how we lead it, and at what pace. On both direction and pace, he claims, we have real control.
That is also why he objects to the name. "There's nothing artificial about artificial intelligence."5:42 He would rather call it augmentative intelligence, a tool that extends human thinking rather than a separate mind. Cieslak's reply was dry: that ship has already sailed.
Cieslak then asked what the risk looks like further out, if doctors become very good at interpreting AI recommendations but less confident about challenging them. Jeelani started his answer with a joke about the Ten Commandments: "Moses did not bring AI down from the mountain." A great deal of what AI is built on is human-generated or human-interpreted knowledge, our record of how we see the world. Human thinking is, in his phrasing, replete with biases and replete with prejudices. Systems trained on that inherit the same shortcomings. His conclusion is the practical one. They are good assistants, but they can make every error a human makes, so "we can't take them as gospel."
What AI did, and did not do, in a conjoined twin separation
The programme had recently followed Jeelani's work separating twins joined at the head. The surgery was successful, and it took both advanced technology and an enormous amount of human skill. That case gave him a concrete example of where AI earned its place.
One task in the most recent separation was narrow and specific. Distractors are surgical devices that push two pieces of bone slowly apart, gradually enough that new bone forms in the gap. Jeelani needed to know how much distraction he could achieve over a few weeks if he placed distractors between two bones. His team had built an AI model to answer exactly that kind of question. They used it, and it was quite accurate.
Then he drew the boundary. For that specific piece of information, AI was wonderful. "The whole package, how you put it together, is still the domain of the doctor." Asking an AI how to separate a set of twins and then relying on what it tells you does not work.
That shape is worth holding on to, because it recurs through the rest of the discussion: a narrow question, a model built for that exact question, and a human still assembling everything around it.
Expertise grows at the edge of your own ability
Cieslak asked whether the risks are different outside fields where people are experts at an elite level. Jeelani's answer inverted the question. The reason he is called an expert, he said, is that he works at the edge of his abilities pretty much every week. Cieslak offered an image back: you are in the swimming pool and your feet are only just touching the bottom.
Many of the operations his team performs have not been done before. They have to push boundaries, find new ways of doing things, and invent new platforms, and each attempt adds a little to what the field knows. Working at the edge of your capabilities, he argued, is what moves you forward. Take that away from the doctors of the future and "they will not progress as we do". His warning about what AI cannot supply is short: "AI is not going to come in and create creativity for them."
Hare brought the same worry from a technology that has already arrived. Robotic surgery is very good for many types of operation, and plenty of surgeons have worried about being replaced. We have not replaced them. We keep human surgeons at the level where they could scrub up and operate today if an emergency demanded it, precisely because, as she put it, "you don't want those skills and experience and confidence to atrophy."
Her generalisation is the part that reaches beyond hospitals. Most of us are not under a surgeon's pressure, but all of us now have two versions of ourselves: the augmented self, with all its tools and technologies, of which AI is probably the most powerful at the moment, and the self that runs on nothing but its own talent and experience. Her worry is about the second one. How mindfully do you train it?
When that second self goes untrained, there is a visible symptom. Asked whether someone can appear more highly skilled than they are because they work alongside AI, Jeelani said it happens all the time. Students arrive with beautiful presentations, "and then you ask them the first question and they fall apart."14:32 What they present is fantastic. Dive into what sits behind it and the substance is not quite there. His summary of the mechanism is old and unglamorous: the brain is a muscle. "You use it or you lose it."
Automation complacency and the muscle you stop training
When Cieslak asked who should have the final say when AI and human judgment disagree, Rinta-Kahila answered that it depends on the context, but that in most cases the decision should sit with the human. He is not dismissive of the tool. He called it a very powerful augmentative aid, especially for people who are already experts and know how to use it mindfully and responsibly.
The problem he pointed to has a name and a long research record behind it: automation complacency, the tendency to stop monitoring a system closely because it almost always works. His example was aviation. The automated systems on a modern aircraft work so well that pilots get used to them working well. The trouble arrives on the rare day when the technology fails, or when the situation is one the technology was never designed to handle.
That is the moment his question lands. "Have you been practicing that muscle?"11:52 Whether it is a cognitive muscle or a physical one, have you actually been training it, and are you ready to react?
He also widened the frame. Medicine is a clear example because it is exactly where we do not want things to go wrong. But modern AI systems are general-purpose technology, able to do many things across many contexts, so the same concern now sits in a great many workplaces that have no equivalent of an operating theatre.
Trust, liability, and who carries the final decision
Jeelani turned the panel's question back on the panel. When, he asked, would you be ready to go to an AI agent and say: I need my gallbladder taken out, I have heard you are very good, I'll put myself in your hands. Both Cieslak and Hare answered without hesitation. Never. "Absolutely never."12:32
His explanation was about trust rather than technical capability. Human interaction, in medicine and beyond it, is fundamentally based on trust or the lack of it. As a children's brain surgeon he meets parents who hand him their sick child, and they do it after reading about him, but the deciding factor is that they look at him, talk to him, and place their trust in a person. Responsibility follows that trust. You can never go back to those parents and say the operation itself went well, but "the AI agent didn't really play ball today". AI can augment the work, he said, "but it cannot take over that role."
The contrast that shows this is not technophobia came a moment later. Asked whether he would get in a self-driving car, he said yes, especially at the end of a late shift when he does not feel like driving home. Driving is a much more repetitive task. The statistics will show in time that self-driving is safer, he said, and they already show it. Surgery, at that level of care and detail, gets a different answer.
From the co-hosts' side of the table came the two variables that decide most of these cases in ordinary work: how much accuracy the task actually needs, and who is liable when it goes wrong. The example was memorable because it is so unheroic. We are probably willing to tolerate more risk in a PowerPoint presentation from a management consultant than in a diagnosis from a doctor, where we want zero risk.
Rinta-Kahila pushed on the same line from the research side. How much risk we accept when handing something to a machine has a lot to do with the cultural and social context we live in, and that context moves. If AI systems become good enough to be shown statistically to produce fewer errors than human surgeons, acceptance may shift with the evidence. His question went beyond surgery: what will we be willing to hand over when machines are, as he put it, better or more trustworthy kindergarten teachers than fallible humans? His own assessment of where we stand today was short. Fortunately, we are not at that stage.
Should children learn the old way before they learn with AI?
The sharpest questions on the panel were about timing rather than technology. Do we need children working with AI right now, or do we need children reading books, given how much trouble literacy rates are already in? Do we want children doing mathematics the old-fashioned way first and graduating into AI later? The same question applies at university, and to junior doctors starting out.
The reason for asking is not nostalgia. Learning something the hard way, the painstaking way, is how you develop a feel for sense-checking an answer, so that you know when AI is wrong. Then the line that sums up the whole episode: "I don't know how you develop that ability if you've never worked without it."17:40
Hare made the same argument at the end of the programme, in something closer to a recommendation. Maybe we do not want this in schools right away, and maybe not for six-year-olds. Maybe it is better at 16 or a bit older, or even at university. Yes, we want to prepare young people for work, but the core skills come first: literacy, numeracy, creativity, constructive thinking, critical thinking. How do we really embed those first? "And that goes for us adults, too."
Read alongside Rinta-Kahila's three definitions, this is the second kind of deskilling and the reason it is treated as the serious one. An expert whose skills erode still has a memory of competence to rebuild towards. Someone who never grappled with the nitty-gritty has no earlier version of themselves to recover.
Organic intelligence and the brain as a prediction engine
Asked what neuroscience says about how the brain develops, Jeelani described a test he runs in lecture theatres. He asks how many people have heard of artificial intelligence, and essentially every hand goes up. Then he asks how many have heard of organic intelligence, the intelligence that runs on biological brains, and at most a handful of people put a hand up. He finds the imbalance fascinating. AI has been in the public domain for perhaps 15 to 20 years. The human brain has been in development rather longer. Cieslak's label for it: intelligence 1.0.
His argument is that many of the questions we are asking about AI, where it will be good and where it needs watching, are answered by lessons already learned from organic intelligence. So he compressed what a brain does into a sentence. We have 86 billion neurons, and only a handful are needed to breathe, eat, and walk around. What do the rest do? "The human brain is a tool of prediction. Period. It's a prediction engine."19:11
The corollary he drew from that is oddly deflating and rather freeing. People who are a little better at predicting the future are the ones we call lucky. "There's nothing godsend about being lucky. It's just their algorithms that little bit better refined." They see one or two steps more clearly than the rest of us. He ended with a proposal that the BBC commission a programme called OI Decoded, organic intelligence decoded, which Cieslak promised to take upstairs.
Set his prediction engine next to Rinta-Kahila's question about practising the muscle, and the deskilling worry gets sharper. If prediction is what the brain does, and if the muscle needs use to stay ready, then every prediction you hand to a machine is one you did not make yourself.
Stickiness: why AI products are built to be hard to give up
A viewer in Columbus, Ohio, asked whether AI is running the Gillette razor strategy: give away the razor, then recover both cost and profit on the blades. Hare thought that was probably pretty accurate, and described what the goal looks like from inside a product team.
What you are looking for when you build a technology or tool is stickiness. Give it away. Let people play with it. Let them see how much it makes their lives better. Then, if you threaten to take it away, what you are hoping is that they will howl and protest, so you can say: that's fine, because we have this lovely subscription model. The aim is for the tool to be embedded in people's lives until they cannot live without it.
What worries her is not the subscription. It is what the dependency does to a person. If you cannot do your work without AI, cannot do school without it, and do not know how to "navigate your emotions without it because it's your therapist, it's your girlfriend, it's your companion", that is where she thinks it gets quite dangerous.
Jeelani agreed on the substance and chose a different word. If we outsource our creativity, something fundamentally human that we start learning at a very young age, "Stephanie used the word dangerous. I'd use the word sad." That, he said, is where we begin to lose our humanity.
His conclusion is not to withdraw. He understands the people who say shut it all down, but thinks the pragmatic reading is that AI is here and will stay whether we like it or not. Our best chance is to understand it, work iteratively with it, and try to keep pace with it or stay ahead of it. Use it for the tasks it is genuinely good at, especially the repetitive ones, to augment our own thinking. His formulation of the correct arrangement fits in one sentence: "It is our plan and AI feeds into it."
Practical habits that keep your own thinking sharp
Rinta-Kahila offered the clearest example on the programme of AI genuinely improving expertise, and the reason it works is as useful as the example. He is an information systems researcher, not an expert in tax law, and tax regulation published on a tax office website is very difficult to interpret. He can upskill himself in tax by talking to an AI that makes it more understandable, then go back and confirm that what it told him is actually right.
Two things make that work. The subject sits outside his core expertise, so there is no professional skill being eroded. And he goes back to the original source instead of trusting the explanation on its own. Note how different that is from Jeelani's case, where the narrow question sat inside his own field rather than outside it. What the two have in common is not how expert the person already is. It is that the question is bounded, and that the human decides what to do with the answer.
Hare's habits work on the same principle from the other end. Your entry point matters, she said. She uses AI on her computer only, has strict rules for herself, and tries not to go to it first. She also says openly that she still gets tired and lazy. The contrast she drew is the person who reaches AI through an Amazon Echo instead: a few of them around the house, the assistant sitting in the ether of home life, children talking to it, and everyone asking it what the weather is, what to cook, what to wear to a party, the fastest way to get across London. "You stop actually just realizing that you could have done that heavy lifting yourself."
She was careful not to turn that into blame. "All of this was pushed out on us really fast. Everybody's experimenting with it. Nobody had best practice." What is different now, she argued, is that studies like Rinta-Kahila's are arriving, so we can be more deliberate.
What follows from the discussion depends on where you sit:
- If you use AI daily: the habits that survive this panel are narrow questions, answers verified against a real source, and a rule about what you do yourself before you ask. Friction is doing a job here, not getting in the way of one.
- If you are training into a profession: the second definition of deskilling is aimed at you. Do the hard version of the work at least once, so you have something of your own to measure the machine's answer against.
- If you manage people: the third definition is a decision, not weather. Simplifying a job so that cheaper staff can fill it has a cost that arrives later, on the day an unusual case turns up and nobody in the room has ever done the work by hand.
- If you are a parent or a teacher: the panel's real disagreement was about sequence, not about whether. Core skills first, AI once there is judgment to apply to its output.
Jeelani's own answer to whether AI can enhance critical thinking rather than replace it was a plain yes, with one condition attached. Point it at the repetitive work, and keep the plan, the questions, and the responsibility on the human side of the table.