What is AI deskilling?
Your team is using AI. Their capability may be quietly declining. Both things can be true at once.
AI deskilling is the process by which repeated AI substitution erodes a specific skill or domain of human competence. It occurs when AI performs tasks that would otherwise develop and sustain human capability, at individual or organisational level.
Deskilling is not a consequence of using AI. It is a consequence of using AI in substitution mode. The same tools, used in augmentation mode, produce the opposite outcome.
Definition from the KAIVANT glossary.
The augmentation-substitution distinction is the most important question in human-AI integration. Most organisations cannot currently answer it. This is what it looks like, why it is almost invisible in the data you are already collecting, and what a structured measurement approach reveals.
When an individual uses AI to handle a piece of work, two things can happen. In the first, AI absorbs execution work and the person invests the freed capacity in more demanding, judgement-intensive work. Output improves. Capability develops. The AI is doing what it is supposed to do: amplifying what the person brings.
In the second, AI handles work that would otherwise have exercised and built the person's reasoning, domain expertise, or judgement. Output still looks fine. The AI completes the task adequately. But the development that would have come from doing that work does not happen. Capability stagnates or declines. The person has not gained time for better work; they have simply removed the work that was making them better.
These two patterns are indistinguishable in short-term output records. They produce identical productivity curves across a quarter. The divergence only becomes visible over months and years, and often only when something changes: the AI fails, the tools are unavailable, or the work requires genuine unassisted judgement in a novel situation.
AI handles the routine, the codifiable, the time-consuming. The person consistently works at the edge of their current competence: harder problems, more novel situations, more genuine judgement. Output improves and capability compounds. The AI makes the person more capable over time.
AI handles not just the routine but the work that was building judgement: the analysis that required synthesis, the decisions that required weighing uncertainty, the writing that required thinking through a problem. The output looks similar. The development is gone. Capability is quietly contracting below the surface of the metrics.
Research on expertise points one way: capability grows through sustained practice at the edge of what a person can currently do. When AI keeps making work easier than a person's skill level, that practice stops. This is not a distant risk. It applies to the way many people already work with AI today.
Three lines of research help explain how AI use can affect this process.
Sparrow, Liu and Wegner (2011) found that people who expected information to stay available on a computer were less likely to remember it, and more likely to remember where to find it. The study was about computers and search, not AI, and a large replication project in 2018 tested one of its experiments and did not reproduce the result. Read it as an early signal, not proof: when people expect a tool to hold knowledge for them, they may put less effort into holding it themselves.
Csikszentmihalyi's flow research links work that is too easy for a person's skill level with boredom and disengagement. The KAIVANT reading of this: when AI keeps removing the hard part of the work, the stretch that builds capability disappears, and a person's capability ceiling can drift down to meet the easier work.
Becker's human capital theory treats skills as capital that people build through investment in learning and practice. Capital that is not maintained loses value. AI does not need to make a person's skills obsolete to deplete them. It only needs to absorb the practice that would have maintained them.
What does recent research show?
Studies from 2025 and 2026 add early evidence. They also suggest that how people use AI matters as much as whether they use it.
In four colonoscopy centres in Poland, experienced doctors found at least one adenoma (a growth that can turn into cancer) in 28.4% of standard colonoscopies without AI before AI assistance was introduced, and in 22.4% after. The study is observational, so other factors may also have played a part. (Budzyń and colleagues, The Lancet Gastroenterology & Hepatology, 2025)
In a randomised trial with 52 mostly junior developers learning a new Python library, the group with AI assistance scored 50% on a follow-up quiz, against 67% for the group without it. Developers who used AI to ask for explanations or conceptual help scored higher than those who handed the coding over. (Shen and Tamkin, Anthropic, 2026)
In a survey of 319 knowledge workers who shared 936 examples from their work, higher confidence in generative AI went with less critical thinking, and higher self-confidence went with more. The data is self-reported and shows an association, not a cause. (Lee and colleagues, Microsoft Research, CHI 2025)
Read together, these studies match the augmentation-substitution distinction. The coding trial shows it most directly: the same tool on the same task went with different learning, depending on how people used it. None of the studies tells a manager which pattern is active in their own team. That needs a measure of capability over time, not of output.
How can people and organisations prevent AI deskilling?
Preventing AI deskilling does not mean using less AI. It means keeping the work that builds skill with people, and checking over time that capability is growing, not only output.
For individuals
- Use AI to understand, not only to produce. Ask for the reasoning, question it, and do the key part yourself. In the coding trial, developers who worked this way scored higher than those who delegated.
- Do some core work without AI, on purpose. Now and then, work unassisted and notice what has become harder.
- Keep choosing problems at the edge of what you can do. That is where capability grows.
For organisations
- Redesign the work, not only the tools. When AI takes over tasks, make sure people still make the judgement calls that matter.
- Protect the practice of early-career staff. They build their foundations through the very work AI can now absorb.
- Measure capability trajectory next to productivity. Output figures alone will not show the difference between the two patterns.
Output metrics measure what was produced, not how it was produced or whether producing it developed the person who did it. A report written with AI assistance and a report written through deep analytical engagement are indistinguishable in any system that measures the report.
Productivity metrics face the same problem. AI-assisted productivity is real productivity. The number of tasks completed, the time taken, the error rate: all of these can improve simultaneously with declining independent capability. The person is more productive in the short term and less capable in the medium term, and nothing in a standard measurement framework will surface that tension until conditions change and the gap becomes visible.
Performance reviews typically lag by months and tend to measure outputs and behaviours, not capability trajectory. By the time a substitution pattern becomes visible in performance data, it has been accumulating for a significant period.
The metrics that tell you AI is working are the same metrics that make capability substitution invisible. Output looks fine. That is exactly what substitution looks like from the outside.
Substitution risk is not uniform across a workforce. It concentrates where specific conditions coincide.
Early and mid-career knowledge workers
Foundational capabilities are built through practice during the formative years of a career. AI that absorbs that practice does not merely slow development. It removes the substrate from which the capability would have grown. The risk is highest in roles where analytical, writing, or judgement skills are the primary development target and AI is used extensively from the outset.
Senior practitioners with high AI delegation
Senior capability stocks are less immediately vulnerable than early-career workers. But they depreciate rather than pause when the judgement work that would maintain them is consistently delegated to AI. The process is slower and the baseline is higher. The risk is cumulative and tends to be noticed late, because the remaining capability is still substantial when the erosion begins.
Organisations that have removed difficulty from the pipeline
At the organisational level, the structural version of deskilling occurs when AI deployment removes the challenging work from workflows without redesigning to ensure human judgement remains engaged at the appropriate level. Organisations can develop a workforce that appears highly productive while aggregate independent capability is in quiet decline. The cross-validation with organisational assessment data is where this becomes visible.
The Kaivant-I is structured around this problem. Its Human Capital Depth axis measures whether independent capabilities are appreciating or depreciating as AI integration increases. Four dimensions are specifically diagnostic for the substitution pattern.
Measures the quality of unassisted reasoning: not as a test of AI-free performance as a goal in itself, but as a periodic calibration of the capability that AI integration is supposed to augment. Declining Judgement Independence is the primary signal that substitution is active rather than augmentation.
Measures the breadth and flexibility of cognitive engagement: whether the work currently draws on an expanding or contracting range of cognitive demands. Contraction of Cognitive Range is a leading indicator: the narrowing that precedes capability decline in specific domains.
Measures whether capability functions at material quality when AI tools are unavailable, and whether that baseline is stable or declining across consecutive assessments. Capability Independence and Resilience is the systemic risk indicator of the Kaivant-I: when it declines over two consecutive periods, the result is flagged, whatever the composite shows. This is not a penalty. It is a structural risk signal.
Measures whether AI use is structured to build judgement or substitute for it: the quality of reflection, deliberate practice patterns, and whether AI-assisted work is producing genuine capability growth. Low Personal Adaptation Architecture with high AI usage is the clearest signal of the substitution pattern: high activity, low development.
The free Kaivant-I Starter covers all nine dimensions and returns a structured read. It does not require an account, takes about 10 minutes, and produces results immediately. Capability Independence and Resilience data is private to the individual by default. This is not a policy choice but an architectural requirement: if others could see it, people would have a reason to perform AI-free work for the measurement instead of engaging genuinely.
Validation status: the method is practice-proven and research-grounded. Statistical validation from KAIVANT's own data grows with every assessment.
Find out whether AI is building or eroding your capability.
The free Kaivant-I Starter takes about 10 minutes. No account required. Your results are immediate and private.
Common questions about AI deskilling
AI deskilling is the process by which repeated AI substitution erodes a specific skill or domain of human competence. It occurs when AI performs tasks that would otherwise develop and sustain human capability, at individual or organisational level. Its critical feature is that it is hard to see: output metrics stay normal while the underlying capability contracts.
Both patterns involve using AI to handle work. In the augmentation pattern, AI absorbs execution work and the freed capacity is reinvested in higher-judgement, more demanding work; output improves and capability develops in tandem. In the substitution pattern, AI handles work that would otherwise have developed the person's judgement, and that developmental work is simply removed. Output may look identical in the short term. Capability trajectories diverge over months and years.
The primary diagnostic question is whether independent capability is stable or declining. Can you perform at material quality when the AI tools are unavailable? Is the range of cognitive challenges in your work expanding or contracting? Are you regularly working at the edge of your current competence, or has AI removed the work that would require that engagement? These are the questions the Kaivant-I measures through its Human Capital Depth axis and the Capability Independence and Resilience dimension.
Output metrics measure what was produced, not how it was produced or whether producing it developed the person who did it. An AI-assisted output and a capability-developing output are indistinguishable in the short-term record. The divergence appears only when conditions change: when a novel problem falls outside what the AI can handle, when tools are unavailable, or when the person is asked to perform unassisted. By then, the capability gap has been accumulating for a significant period.
Risk is highest where three conditions coincide: high AI usage, early or mid-career stage where foundational capabilities are still being built through practice, and no deliberate effort to maintain capability through work that requires genuine unassisted engagement. Senior practitioners with established capability stocks face a different risk: slower but cumulative depreciation as AI absorbs the judgement work that would otherwise maintain those stocks. Both patterns are measurable through the Kaivant-I.
Organisations can develop a structural form of AI deskilling when they deploy AI extensively without maintaining the human judgement capacity to oversee it. This appears as high Human Judgement Utilisation claims that are contradicted by actual individual capability readings. The organisation believes its people are doing judgement work; their independent judgement capacity is declining; the gap between the organisational assumption and individual reality is a systemic risk that accumulates silently. The cross-validation between organisational and individual instrument scores is where this becomes detectable.
Early evidence points one way. In four colonoscopy centres in Poland, doctors found at least one adenoma in 28.4% of colonoscopies without AI before AI assistance was introduced, and in 22.4% after (an observational study). In a randomised trial, developers who learned a new coding library with AI scored 50% on a follow-up quiz, against 67% without AI. In the same trial, developers who used AI to ask for explanations scored higher than those who delegated the work.
Use AI to understand, not only to produce: ask for the reasoning and do the key part yourself. Do some core work without AI on purpose, and keep taking on problems at the edge of your ability. In organisations, redesign work so people still make the judgement calls that matter, protect the practice of early-career staff, and measure capability over time next to productivity.