Post
JA EN

The More You Use It, the Less You Can Do Without It — Empirical Research on the AI Deskilling Paradox

Posted
By
18 min read
The More You Use It, the Less You Can Do Without It — Empirical Research on the AI Deskilling Paradox
  • Target audience: Software engineers who use AI tools on a daily basis
  • Prerequisites: Basic experience with AI coding tools such as GitHub Copilot and ChatGPT
  • Reading time: 15 minutes

Summary

AI tools are designed to augment human capabilities. Yet recent research suggests that AI assistance may not augment but rather atrophy human skills. A study published in The Lancet found that after AI-assisted endoscopy was introduced at medical facilities, physicians’ lesion detection rates during non-AI examinations dropped from 28.4% to 22.4%. An analysis published by ACM identifies a positive feedback loop in which AI adoption leads to skill degradation and deepening dependence. This article examines the mechanisms, evidence, and implications of the AI deskilling paradox for engineers, drawing on multiple peer-reviewed studies.

What Is the AI Deskilling Paradox?

“Deskilling” refers to the phenomenon in which the introduction of technology renders workers’ existing skills unnecessary and causes them to atrophy. The same dynamic that occurred when factory automation made skilled craftspeople’s expertise obsolete is now being replicated between AI and knowledge workers.

However, AI deskilling differs qualitatively from traditional automation. Factory automation replaced physical tasks, but AI replaces cognitive tasks — judgment, analysis, and problem structuring. And cognitive skills have the property of “use it or lose it”1.

An analysis published in Communications of the ACM clearly illustrates the structure of this paradox1:

flowchart TB
    A["AI makes work<br>more efficient"] --> B["Fewer opportunities<br>to use foundational skills"]
    B --> C["Skills atrophy"]
    C --> D["Dependence on AI<br>deepens further"]
    D --> A

    style A fill:#f9f,stroke:#333
    style C fill:#fbb,stroke:#333

What makes this feedback loop problematic is that the first step is something positive (efficiency gains). Efficiency itself is desirable. But as a consequence of that efficiency, skills atrophy, and atrophy breeds further dependence. Once you enter this loop, it becomes difficult to remove AI assistance.

Historical Context: The Recurring Pattern of Technology and Skill Degradation

AI deskilling is the latest installment of a recurring pattern in the history of technology.

TechnologyConvenienceDegraded SkillResearch Evidence
CalculatorFaster computationMental arithmeticWidely reported
GPSAutomated navigationSpatial cognition & sense of directionDahmani & Bohbot (2020), etc.
AutocorrectAutomatic spelling correctionSpelling abilityNoted in educational research
AIAutomation of cognitive tasksCritical thinking & problem analysisGerlich (2025), etc.

The common pattern is clear: we gradually lose foundational abilities in exchange for convenience. However, there are important differences in this analogy.

Calculators and GPS replaced specific skills. Even if mental arithmetic ability declined, other aspects of mathematical thinking were unaffected. But AI has the potential to replace cognitive processes in general — analysis, judgment, problem structuring, and critical evaluation. The scope of impact is qualitatively different2. The risks of this “outsourcing of cognitive processes” are analyzed in detail from a mechanistic perspective in the article on cognitive offloading. This article focuses on the empirical data and structural factors behind skill degradation.

Lancet Gastroenterology & Hepatology Study: Deskilling Demonstrated in Clinical Practice

Study Overview

The moment the AI deskilling paradox was elevated from “theoretical concern” to “demonstrated problem” came with a study published in The Lancet Gastroenterology & Hepatology in 20253. This study examined the impact of AI use on subsequent physician performance, focusing on AI-assisted colonoscopy (Computer-Aided Detection, CADe).

Study design:

ItemDetails
SubjectsEndoscopists at 4 facilities in Poland
AI assistanceCADe (Computer-Aided Detection) system
DesignRetrospective comparison before and after AI introduction
MeasurementDetection rates in non-AI examinations before vs. after AI introduction
Primary outcomeAdenoma Detection Rate (ADR)

Results

PeriodAdenoma Detection Rate (ADR)
Before AI introduction (non-AI exams)28.4%
After AI introduction (non-AI exams)22.4%
Change-6 percentage point decline

In examinations performed without AI after AI introduction, the adenoma detection rate dropped from 28.4% to 22.4% 3. Even in an environment where AI-assisted and non-AI examinations alternated, the physicians’ ability to “see with their own eyes” during non-AI examinations had declined.

Why Did Detection Rates Decline?

The most plausible interpretation of these results is reallocation of attentional resources. Physicians who had adapted to an environment where AI pointed out lesions in alternating examinations may have unconsciously reduced the cognitive attention they devoted to independent searching. As a result, they could no longer maintain the same level of attention during non-AI examinations.

This is closer to “deactivation of ability” than “loss of ability.” But the practical impact is the same — patient risk increases in situations where AI is unavailable.

Limitations of This Study

Several caveats apply to the interpretation of this study. First, it is unclear how long after AI introduction the deskilling effect takes hold, or how long recovery would take. ADR varies considerably between facilities and individuals, and whether a 6-percentage-point decline is clinically significant in all contexts is debatable. Furthermore, this study focused on a specific medical procedure, and caution is needed when generalizing to cognitive tasks like software development.

The ACM’s Analysis of the Deskilling Paradox Structure

Three Skill Domains Affected

An analysis published in Communications of the ACM, based on research from Microsoft Research and Carnegie Mellon University, identifies three skill domains affected by AI deskilling1:

1. Loss of Foundational Knowledge

When AI provides instant answers, the motivation to learn a field’s foundational knowledge independently declines. In the context of programming, the perceived need to memorize things like basic language syntax, standard library functions, and data structure characteristics — things that “you can just look up” — diminishes.

2. Deterioration of Problem Analysis and Diagnosis Skills

When you repeatedly paste error messages into AI and receive fix suggestions, you fail to develop the ability to analyze root causes independently. Debugging is a cycle of hypothesis generation and testing, and going through that cycle is itself the source of the skill.

3. Decline in Social Interaction Skills

Asking AI questions reduces opportunities to ask colleagues, making it harder to develop communication skills for technical discussions and code reviews. This is a particularly overlooked risk for junior engineers.

Concrete Scenarios in Engineering

The following are deskilling scenarios in software development inferred from the ACM analysis and the Lancet study findings. Note that these are not based on empirical studies directly targeting software development, but rather inferences drawn from medical and cognitive science findings.

SkillWith AI dependencyWhat may happen when AI is unavailable
DebuggingPaste errors into AI for fixesForget how to read stack traces
DesignHave AI suggest architectureAbility to evaluate trade-offs declines
Code readingAsk AI “What does this code mean?”Lose ability to trace complex code independently
Test designHave AI generate test casesAbility to discover edge cases atrophies

Findings from Education Research: Three Adaptation Patterns

Deskilling Is Not Inevitable

A 2025 study published on ScienceDirect classified student AI adaptation into three patterns, demonstrating that deskilling is not the only outcome4:

1. Deskilling (Skill Degradation)

  • Existing skills weaken due to AI dependence
  • An attitude of “I don’t need to learn this because AI can do it”
  • Passive usage patterns

2. Reskilling (Acquiring New Skills)

  • Acquiring new skills related to AI use
  • Prompt engineering, evaluating AI output, workflow design
  • Existing skills degrade but are supplemented by new skills

3. Upskilling (Skill Enhancement)

  • Using AI as a lever to strengthen existing skills
  • Critically evaluating AI output to deepen one’s own understanding
  • Using AI as a “teacher”

What Determines the Fork in the Road

The most significant factor the research identified was individual learning goals 4. Students with performance goals — “I want to finish this task quickly” — tended toward deskilling, while students with mastery goals — “I want to develop skills” — tended toward upskilling.

This is suggestive in an engineering context as well. Using Copilot as “a tool to write code faster” versus “a tool to improve the quality of my code” may have different long-term effects on skill development.

flowchart TB
    AI["AI Tool Usage"]
    AI --> G1["Performance Goal<br>\"I want to finish quickly\""]
    AI --> G2["Mastery Goal<br>\"I want to develop skills\""]

    G1 --> D["Deskilling<br>(Skill Degradation)"]
    G2 --> U["Upskilling<br>(Skill Enhancement)"]

    G1 --> R["Reskilling<br>(New Skill Acquisition)"]

    classDef negative stroke:#d29922,stroke-width:3px
    classDef positive stroke:#2ea44f,stroke-width:3px
    class D negative
    class U positive

Insights from Anthropic’s Economic Index Report

Anthropic’s published economic index report analyzes deskilling trends from actual Claude usage data5.

The key concept to note in this report is “net deskilling effect.” When AI reduces demand for some skills (deskilling) while increasing demand for others (upskilling), the net impact must be evaluated as the difference between the two.

The report’s analysis shows that AI substitution is progressing particularly in routine cognitive tasks5. As the time spent on these tasks is replaced by AI, the skills associated with those tasks see reduced usage frequency.

On the other hand, new skill demands are also emerging — AI output quality management, prompt design, and AI workflow construction. The problem is that the skills that atrophy and the skills newly demanded are not necessarily distributed equally across the same individuals.

Deskilling as a Structural Problem

A 2025 paper published in AI & Society frames AI deskilling not as individual negligence but as a structural problem 2.

This perspective is important. If you dismiss it as “relying too much on AI is an individual problem,” the countermeasure becomes the platitude “think more for yourself.” But deskilling arises from structures where organizations adopt AI, measure KPIs by efficiency gains, and view time spent working without AI as “unproductive.”

Consider a concrete example in engineering organizations:

  • Using AI for code generation becomes the team standard
  • Deadlines are shortened with AI-generated code (efficiency KPIs are met)
  • But opportunities to write code without AI decrease
  • Months later, productivity in situations where AI is unavailable (security constraints, offline environments, proprietary frameworks AI cannot handle) has declined compared to before

This structure cannot be dismissed as “the individual’s fault.” Skill maintenance must be designed into the system at the organizational level.

Future Challenges Highlighted by a Medical Review Study

A mixed-methods review published in Artificial Intelligence Review systematically organized research on AI-induced deskilling in medicine and presented a future research agenda6.

The key challenges this review identifies are as follows:

  1. Long-term impacts remain unclear: There are virtually no studies tracking skill changes over months to years
  2. Recoverability is unknown: After deskilling occurs, the extent and conditions under which recovery is possible have not been studied
  3. Individual difference factors: Even with the same AI assistance, there are large individual differences in the degree of deskilling. Identifying the factors (years of experience, metacognitive ability, learning orientation, etc.) is urgent
  4. Organizational countermeasures are unevaluated: The effectiveness of measures such as ensuring training time without AI and conducting regular skill assessments has not been verified

Discussion: How to Make AI and Skills Coexist

“Change How You Use AI” Rather Than “Don’t Use AI”

The intuitive countermeasure to deskilling is “set aside time without AI.” The Lancet study also proposes alternating between “AI-assisted days” and “non-AI days”3.

However, given the reality that AI is becoming a standard tool, is the prescription of “periodically removing AI” sustainable in the long term? After calculators became widespread, being told to “do your work with mental arithmetic for a few hours each week” would not have been realistic.

More recent evidence suggests that the core of the problem lies not in whether AI is present, but in the degree of cognitive engagement. Shen & Tamkin (2025) conducted a randomized controlled experiment with 51 professional developers7. The AI-using group overall scored 17% lower on comprehension tests. However, a clear fork existed within AI users.

The research identified six AI interaction patterns, which fell broadly into two groups:

Pattern GroupExamplesComprehension Test Score
High-engagement patternsGenerate-then-understand, hybrid explanation, conceptual questioning65-86%
Low-engagement patternsWholesale delegation, incremental dependence, iterative debugging24-39%

Developers who used high-engagement patterns with AI achieved scores equal to or higher than the group that did not use AI at all (67%). In other words, it is possible to maintain skills while using AI, and the key lies in how you use it7.

A PNAS study by Bastani et al. (2025) confirmed the same structure8. In a large-scale experiment with approximately 1,000 high school students, the group that used GPT-4 without restrictions showed 17% lower performance after AI was removed compared to the AI-inexperienced group. However, in the AI group with educational guardrails that “provided hints but did not give answers,” this negative effect was substantially mitigated.

What these studies demonstrate is that the essence of deskilling lies not in “using AI” but in “delegating the entire cognitive process to AI.”

Which Skills to Maintain and Which to Let Evolve

The premise that “all skills should be maintained” is worth questioning. Just as most people are untroubled by their declining mental arithmetic or GPS navigation abilities, there are situations where the cost of maintaining skills for tasks AI can handle with high reliability outweighs the benefit.

However, the “calculator analogy” has important limitations. Calculators are deterministic and do not make mistakes, but LLMs hallucinate. Calculators replaced an isolated computational skill, but AI affects interconnected cognitive skills such as analysis, judgment, and structuring. The fact that skill degradation is hard to notice just from the appearance of the output is also qualitatively different from calculators.

A more practical framework is to divide skills into “execution skills” and “evaluation skills”:

  • Execution skills (can be delegated to AI): Writing boilerplate, routine CRUD operations, format conversion. AI reliability is high in these areas, and the cost of maintaining these skills may not be justified
  • Evaluation skills (should be retained by humans): Problem structuring, architectural decision-making, security risk assessment, and the ability to determine whether AI output is “correct.” AI reliability is unstable in these areas, and these skills are also prerequisites for using AI effectively

An important caveat here: evaluation skills cannot be formed without execution experience. The ability to analyze root causes in debugging grows from the experience of debugging yourself. The judgment to assess design trade-offs is learned from design failures. Therefore, especially in the early stages of a career, execution skill experience is indispensable for building the foundation of evaluation skills. The notion that “execution can be left to AI” only holds for those who already have a foundation in evaluation skills.

Metacognitive Defense — Changing How You Use AI Through How You Use It

As the education research showed, the key that determines whether you experience deskilling or upskilling is setting learning goals 4. This is closely related to metacognition — the ability to monitor your own thought processes and recognize “am I thinking right now, or am I letting AI think for me?”

Translating the “high-engagement patterns” identified by Shen & Tamkin into engineering practice:

  • Generate-then-understand pattern: After having AI generate code, analyze yourself “why this implementation?” and “how does it compare to other approaches?”
  • Hybrid explanation pattern: While implementing yourself, have AI critique “what are the problems with this approach?”
  • Conceptual questioning pattern: Ask AI not “what should I write?” but “why is this design good?”

What these share in common is whether your own cognitive process is active after receiving AI output. The habit of asking yourself not just “let AI generate it and I’m done” but rather “what did I learn from AI’s output?” and “did my understanding deepen?” becomes a bulwark against deskilling. This aligns with the finding that metacognition determines the success or failure of AI use.

Conclusion

Here is a summary of the research findings on the AI deskilling paradox.

Empirically demonstrated findings:

  • At facilities where AI-assisted endoscopy was introduced, detection rates in non-AI examinations dropped from 28.4% to 22.4%3
  • AI adoption can create a positive feedback loop of efficiency gains, reduced skill usage opportunities, skill degradation, and deepened dependence1
  • Student AI adaptation falls into three patterns — deskilling, reskilling, and upskilling — with learning goals determining the fork4
  • Whether skills are maintained while using AI depends on the degree of cognitive engagement. With “high-engagement patterns,” skills were maintained at levels equivalent to the non-AI group7

Suggestive findings requiring further verification:

  • Similar deskilling effects in software development (direct empirical studies are lacking)
  • Generalizability of educational guardrails’ deskilling mitigation effects to other domains8
  • Recoverability of deskilling and the conditions under which it occurs

Unresolved questions:

  • Criteria for which skills to maintain and which skill degradation to accept
  • Methods for evaluating the net deskilling effect (skills that atrophy vs. newly acquired skills)
  • Long-term (multi-year) skill change trajectories
  • How to develop “evaluation skills” in environments where AI is always available

Deskilling is not “negligence” — it is a “structural problem”2. Yet the countermeasure is not the simple prescription of “set aside time without AI,” but rather designing systems that maintain cognitive engagement while using AI.

For more on this topic, see the following related articles:

References

References are listed below in the order they are cited in the text.

  1. The AI Deskilling Paradox - Communications of the ACM (2025). Analysis based on research from Microsoft Research and Carnegie Mellon University. Structurally organizes the positive feedback loop and impacts on three skill domains resulting from AI adoption. [Reliability: High] ↩︎ ↩︎2 ↩︎3 ↩︎4

  2. AI deskilling is a structural problem - AI & Society, Springer (2025). Frames AI deskilling as a structural problem rather than individual negligence. Peer-reviewed. [Reliability: High] ↩︎ ↩︎2 ↩︎3

  3. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study - The Lancet Gastroenterology & Hepatology, Vol. 10, Issue 10, pp. 896-903 (2025). DOI: 10.1016/S2468-1253(25)00133-5. Retrospective comparison at 4 facilities in Poland. ADR in non-AI examinations dropped from 28.4% to 22.4% after AI introduction. Empirical study of AI deskilling based on real clinical data. Peer-reviewed. [Reliability: High] ↩︎ ↩︎2 ↩︎3 ↩︎4

  4. Deskilling, reskilling, or upskilling? Unpacking the pathways of student adaptation to generative artificial intelligence - Information & Management (2025). Classifies student AI adaptation into three patterns. Individual learning goals determine the fork. Peer-reviewed. [Reliability: High] ↩︎ ↩︎2 ↩︎3 ↩︎4

  5. New building blocks for understanding AI use - Anthropic Economic Index (2025). Analyzes the possibility of “net deskilling effects” from Claude usage data. Industry report. [Reliability: Medium-High] ↩︎ ↩︎2

  6. AI-induced Deskilling in Medicine: A Mixed-Method Review and Research Agenda - Artificial Intelligence Review, Springer (2025). Mixed-methods review of AI-induced deskilling in medicine. Presents a future research agenda. Peer-reviewed. [Reliability: High] ↩︎

  7. How AI Impacts Skill Formation - Shen, E. & Tamkin, A. Anthropic (2025). Randomized controlled experiment with 51 professional developers. Identified six AI interaction patterns and showed that “high-engagement patterns” maintained skills at levels equivalent to the non-AI group. Preprint. [Reliability: Medium-High] ↩︎ ↩︎2 ↩︎3

  8. Generative AI without guardrails can harm learning: Evidence from high school mathematics - Bastani, H., et al. Proceedings of the National Academy of Sciences (PNAS) (2025). Large-scale field experiment with approximately 1,000 high school students. The unrestricted AI group showed 17% lower scores after AI removal, but the negative impact was mitigated in the AI group with educational guardrails. Peer-reviewed. [Reliability: High] ↩︎ ↩︎2

AI・Technology
AI Deskilling Skill-Degradation Medical-AI Software-Development
This post is licensed under CC BY 4.0 by the author.