Post
JA EN

The Truth Behind Experts Who Seem to 'Blindly Delegate' to AI: The Speed-Quality Paradox

Posted
By
15 min read
The Truth Behind Experts Who Seem to 'Blindly Delegate' to AI: The Speed-Quality Paradox
  • Target Audience: Software engineers, IT professionals interested in AI adoption
  • Prerequisites: Basic experience with AI tools such as GitHub Copilot, ChatGPT, Claude, etc.
  • Reading Time: 20 minutes

Overview

“Why can that person produce such high-quality work when they seem to blindly delegate everything to AI without thinking?”

The answer to this question is counterintuitive. What recent research has revealed is that experts actually become “slower”. Yet at the same time, quality improves. In other words, experts are not “blindly delegating”—they are converting speed into quality.

This article examines how experts interact with AI based on recent research from METR (2025), Qodo (2025), Fügener et al. (2022), the Dreyfus model, and Kahneman & Klein’s research on expert intuition. We present evidence-based insights into what “appears to be blind delegation” actually is and how to develop this capability.

The Phenomenon: The Expert Paradox

Surprising Research Results—Experts Get “Slower”

In July 2025, METR published a study involving 16 experienced open-source developers1. The subjects were experts who had contributed for multiple years to repositories with an average of over 22,000 stars and more than 1 million lines of code.

The results were shocking:

“When developers used AI tools, tasks took 19% longer—a significant slowdown contrary to developer beliefs and expert predictions”1

Even more surprising was the gap with developers’ own perceptions:

“Developers expected to be 24% faster with AI. And even after actually becoming slower, they believed they had become 20% faster”1

Behavioral Pattern Comparison: Senior vs. Junior

Qodo’s 2025 survey discovered interesting patterns between senior and junior developers2:

 Senior DevelopersJunior Developers
Quality improvement perception60% (highest)Low
Confidence in AI output22% (lowest)High
Perceiving context issues52%41%

Another study (250 participants) confirmed similar patterns3:

 Senior DevelopersJunior Developers
AI suggestion review timeAverage 4.3 minAverage 1.2 min
Productivity change10-15% decrease30-40% increase

The Core of the Paradox

Here lies the paradox:

  • Seniors become slower (10-19%)
  • Seniors don’t trust AI output (22% confidence)
  • Yet seniors have the greatest quality improvement (60%)

Researchers point out:

“True expertise isn’t just knowledge—it’s intuition, pattern recognition, and deeply internalized workflows. Tools that interrupt these face resistance—and that resistance often reflects not mere stubbornness, but genuine wisdom”3

In other words, experts produce “high quality” work not because they use AI efficiently, but because they rigorously evaluate AI output.

Why It Looks Like “Blind Delegation”

A study analyzing 168,000 AI suggestions discovered an interesting pattern4:

“The most common pattern was programmers writing new code features and validating displayed suggestions in cycles. When writing new features, programmers don’t pause for suggestion verification—they continue writing while rejecting”

In other words, experts are not “accepting without thinking”—they are “making instant judgments to reject while continuing work”. This “judgment that doesn’t appear to involve thinking” is what looks like “blind delegation.”

Cognitive Automation—The Mechanism of Invisible Judgment

Research on automaticity reveals the mechanism behind this “invisible judgment”5:

“Operations that initially are slow, sequential, and require conscious attention become, with practice, fast, unintentional, and capable of being executed in parallel with other processes”

Reasons why expert judgment appears “invisible”:

  1. System 1 judgment: Intuitively judging “this is wrong” without reaching consciousness
  2. Reduced cognitive load: Automated judgment consumes almost no cognitive resources
  3. Parallel processing: Can think about the next code while making rejection judgments
  4. Difficult to verbalize: Even they struggle to explain “why they rejected”

From the outside it looks like “blind delegation without thinking,” but in reality massive rejection judgments are being processed at high speed beneath the surface of consciousness.

However, there’s an important caveat. This “speed” may be an illusion. As the METR study showed, experts actually become 19% slower than without AI. It “appears fast” because conscious processing isn’t visible, not because overall productivity is high.

Theoretical Framework

The Dreyfus Model: Five Stages of Skill Acquisition

In 1980, UC Berkeley researchers Stuart Dreyfus and Hubert Dreyfus developed a model of skill acquisition at the request of the U.S. Air Force for pilot training improvement6.

flowchart TB
    N["1. Novice: Follows rules"]
    AB["2. Advanced Beginner: Applies rules situationally"]
    C["3. Competent: Plans and feels responsible"]
    P["4. Proficient: Judges intuitively"]
    E["5. Expert: Acts unconsciously, automatically"]

    N --> AB --> C --> P --> E

    style E stroke:#2ea44f,stroke-width:3px

The core of the Dreyfus model is the discovery that with expertise comes a transition from analysis to intuition, from conscious to automatic processing6:

“An expert with vast experience in a variety of situations acts intuitively without reflective decision-making. As the Dreyfus brothers stated in ‘Mind Over Machine’: ‘When things are proceeding normally, experts don’t solve problems and don’t make decisions; they do what normally works’”

This is why it looks like “blind delegation.” Expert judgment is made without conscious analysis. From the outside it appears they’re “not thinking,” but in reality instantaneous pattern matching based on vast experience is occurring.

Metaknowledge: The Ability to “Know What You Know”

Fügener et al. (2022)’s study in Information Systems Research revealed a fundamental problem in human-AI collaboration7:

“Human performance was impaired by lack of metaknowledge—humans couldn’t accurately assess their own abilities, resulting in inappropriate delegation decisions”

Metaknowledge is the ability to “know what you know and what you don’t know.” The key findings of this research:

  1. AI is better at delegation than humans: When AI delegated to humans, overall performance improved even when delegating to low-performing humans
  2. Humans are poor at self-assessment: Humans were poor at delegating to AI and couldn’t benefit from AI
  3. The cause is lack of metaknowledge: Not algorithm aversion, but inability to accurately grasp one’s own abilities was the cause

“Decision-makers with strong metaknowledge can delegate appropriately because they know whether their answer is correct. With insufficient metaknowledge, one might be confident in wrong answers and doubt correct ones”7

In other words, experts can “delegate” because they can instantly judge “this is safe to leave to AI” or “I should do this myself.”

Tacit Knowledge: Knowledge That Can’t Be Put Into Words

Philosopher Michael Polanyi presented the famous proposition in the 1950s that “we can know more than we can tell”8.

Tacit knowledge is knowledge that is difficult to verbalize:

  • How to ride a bicycle
  • An experienced programmer’s intuition that “this code is suspicious”
  • A veteran doctor’s ability to judge “this patient is serious” at a glance

Cognitive Task Analysis (CTA) research has revealed a surprising fact9:

“Structured elicitation surfaces critical tacit knowledge that experts often omit—40-70% of critical steps and cues are omitted when teaching novices without CTA”

In other words, experts don’t fully recognize what they’re doing themselves. This is another reason they “appear to not be thinking.” Highly sophisticated judgment is occurring, but it doesn’t surface to consciousness.

System 1 and System 2: Dual Process Theory

The dual process theory, made famous by Daniel Kahneman’s book Thinking, Fast and Slow, also explains this phenomenon10:

System 1 (Fast Thinking)System 2 (Slow Thinking)
Automatic, intuitiveConscious, analytical
EffortlessRequires effort
Pattern recognitionLogical reasoning
Expert’s normal modeNovice’s normal mode

In medical education research11:

“Diagnostic decision-making is performed through a combination of System 1 (intuitive or pattern recognition) and System 2 (analytical) thinking. Complex cognitive operations transition from System 2 to System 1 as expertise and skill are acquired (becoming more automatic)—pattern matching replaces effortful reasoning”

Experts can “instantly” reject AI suggestions because System 1 is judging “this is wrong.”

Kahneman & Klein: When Can Expert Intuition Be Trusted?

This raises an important question. Is expert intuition always trustworthy?

In 2009, Daniel Kahneman (the “data-driven” skeptic of intuition) and Gary Klein (the “intuition-trusting” researcher) published a joint paper12. Their conclusion clarified the conditions under which expert intuition can be trusted:

“For intuitive expertise to develop, (1) an environment with sufficiently predictable regularities, and (2) opportunities to learn those regularities through prolonged practice are necessary”12

That is:

Environments where intuition is reliableEnvironments where intuition is dangerous
Patterns recur repeatedlySituations differ each time
Feedback is immediateResults are delayed
Sufficient experience can be accumulatedExperience is limited
Examples: surgeons, firefighters, chessExamples: stock prediction, political forecasting

And an important warning:

“Subjective confidence is not a reliable indicator of judgment accuracy”12

Klein offers practical advice:

“‘I can act on my intuition, no worries’ is a dangerous attitude. Treat intuition as an important data point while consciously and deliberately evaluating it to see if it makes sense in this context”12

These insights apply to expert judgment in AI use:

  1. Coding is a relatively “high validity environment”: Patterns recur, errors become apparent relatively quickly
  2. But intuition doesn’t work well with new technologies/domains: Due to lack of experience
  3. Watch out for the confidence-accuracy gap: As the METR study showed, even experts misjudge their own performance

What Research Shows About Experts and AI

Why Can Experts Use AI Effectively?

Imundo et al.’s research explains why experts can effectively leverage AI13:

“Experts have well-structured content knowledge and procedural knowledge within the domain, making them suited to leverage GenAI for higher-order decision-making. This expertise enables them to formulate effective prompts for generating relevant answers based on accurate domain-specific terminology and knowledge structures”

In contrast, novices:

“Novice writers may bypass critical writing processes such as constructing logical arguments and understanding the subject matter, potentially impairing long-term skill development”13

The Existence of the “Danger Zone”

Interestingly, there’s a “danger zone” in AI use14:

“When ChatGPT fabricates plausible analyses in domains the user doesn’t understand, the failure goes undetected. The need for new ‘metacognitive skills’ is recognized, but that’s precisely what novices in the danger zone lack”

“Senior practitioners don’t just know the answers—they understand the reasoning behind the answers, the historical context that shaped current practices, and subtle indicators that standard approaches won’t work”14

The Metacognition Paradox: Higher AI Proficiency, Worse Self-Assessment

Fernandes et al.’s study (N=246) discovered a surprising paradox15:

“Participants using AI to solve LSAT (Law School Admission Test) logical reasoning problems improved task performance by 3 points while overestimating their own task performance by 4 points”

Furthermore:

“Interestingly, higher AI proficiency correlated with lower metacognitive accuracy—those with more technical knowledge about AI were more confident but no better at accurately judging their own performance”

This suggests that “knowing about AI” and “using AI effectively” are different things.

Practical Implications for Engineers

Visualizing Expert Judgment Processes

The following diagram shows the difference in how experts and novices respond to AI suggestions:

flowchart LR
    subgraph Expert["Expert Processing"]
        direction TB
        E1["Receive AI suggestion"]
        E2{"System 1<br>Intuitive judgment"}
        E3["Immediate rejection<br>Continue working"]
        E4["Immediate acceptance<br>Continue working"]
        E5["Activate System 2<br>Detailed examination"]

        E1 --> E2
        E2 -->|"Something's off"| E3
        E2 -->|"No problem"| E4
        E2 -->|"Judgment pending"| E5
    end

    subgraph Novice["Novice Processing"]
        direction TB
        N1["Receive AI suggestion"]
        N2["System 2<br>Conscious analysis"]
        N3{"Understand?"}
        N4["Accept<br>(even if not fully understood)"]
        N5["Reject<br>(vague reasoning)"]

        N1 --> N2
        N2 --> N3
        N3 -->|"Sort of"| N4
        N3 -->|"Don't know"| N5
    end

    style E2 stroke:#2ea44f,stroke-width:3px
    style N3 stroke:#d29922,stroke-width:3px

How to Develop This Capability

Research-suggested development approaches:

1. Deliberate Practice

According to the Dreyfus model, transition to System 1 occurs through extensive experience6. Not simple repetition, but deliberate practice in various situations is needed.

Practical approaches:

  • Intentionally allocate time to code without AI
  • When encountering errors, form hypotheses before asking AI
  • Before accepting AI suggestions, explain to yourself “why this is correct”

2. Strengthening Metacognition

Metaknowledge research shows the importance of self-assessment ability7.

Practical approaches:

  • Consciously judge whether “I should do this myself or leave it to AI”
  • Build the habit of verbalizing the basis for judgment
  • Don’t be ashamed to admit “I don’t know”

3. Verbalizing Tacit Knowledge

CTA research shows the value of making tacit knowledge conscious9.

Practical approaches:

  • Verbalize “why this felt suspicious” in code reviews
  • Think aloud during pair programming
  • Trust “intuition” while verifying reasons afterward

4. Recognizing the “Danger Zone”

In areas where you’re a novice, AI output needs particularly careful evaluation14.

5. Don’t Overconfidently Trust Intuition (Kahneman & Klein’s Lesson)

Recognize conditions where intuition is reliable, and be aware of the confidence-accuracy gap12.

Practical approaches:

  • Question intuition with new technologies/domains
  • Don’t take the subjective feeling of “getting faster” at face value (lesson from METR study)
  • Perform conscious verification after intuitive judgment
  • Always verify AI output with multiple sources in unfamiliar domains
  • When in doubt, consult experts (humans)

Practical Examples with GitHub Copilot / Cursor

Expert approach:

1
2
3
4

1. First think about design (without AI)
2. View AI suggestions with implementation ideas in mind
3. When suggestions differ from your ideas, consider why
4. If it "feels wrong," immediately reject and continue writing

Novice approach (to avoid):

1
2
3
4

1. Just start writing code
2. Accept AI suggestions as-is
3. It works, so OK
4. Understand it later (or not)

Summary

The Expert Paradox: Slower but Higher Quality

From this article’s analysis, the following has become clear about expert AI use:

  1. Experts actually “get slower” (METR study: 19%, other studies: 10-15%)
  2. But quality improves (Qodo study: 60% perceive quality improvement)
  3. “Appears to be blind delegation” because judgment is instant and doesn’t reach consciousness

In other words, experts are not “blindly delegating”—they are converting speed into quality.

Why Experts Can Ensure Quality

  1. Unconscious competence (Dreyfus model): Instant judgment via System 1 through vast experience
  2. Metaknowledge: Can accurately judge “should I do this myself or leave it to AI”
  3. Tacit knowledge: Non-verbalizable knowledge supports intuitive quality assessment
  4. High rejection rate: Maintains quality by rejecting much of AI output

Important Caveats

As Kahneman & Klein’s research shows, expert intuition is not always trustworthy. Intuition is effective when:

  • In environments with predictable patterns
  • When there are sufficient learning opportunities

And subjective confidence is not a reliable indicator of accuracy. As the METR study showed, even experts misperceive that they’ve “gotten faster.”

Implications for Novices

Imitating experts’ “blind delegation” is meaningless because they are “not blindly delegating.” The path to expertise:

  • Deliberate practice: Build foundations without AI
  • Strengthen metacognition: Accurately understand the limits of your abilities
  • Accumulate tacit knowledge: Develop intuition through diverse experiences
  • Verify intuition: Don’t over-trust intuition in new domains

Ultimately, the difference between experts and novices is not “how much they use AI,” but “whether they can appropriately reject AI output.” And that capability depends on expertise cultivated outside of AI.

On citation accuracy: The research cited in this article has been verified through the following methods:

  • Confirmation in academic databases (Google Scholar, PubMed, etc.)
  • Verification of paper information on official journal websites
  • Cross-verification through multiple independent sources

Check out other articles related to this theme:

References

References corresponding to citation numbers in the text are listed in numerical order.

Additional References (not numbered in text)

  1. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity - METR (2025). [Reliability: High] ↩︎ ↩︎2 ↩︎3

  2. State of AI Code Quality in 2025 - Qodo (2025). [Reliability: Medium-High] ↩︎

  3. Why Sr. Devs Are Actually Less Productive with AI Copilot - DZone (2024). [Reliability: Medium-High] ↩︎ ↩︎2

  4. Reading Between the Lines: Modeling User Behavior and Costs in AI-Assisted Programming - Mozannar, H. et al. (2022). arXiv preprint. [Reliability: Medium-High] ↩︎

  5. Controlled and automatic human information processing: I. Detection, search, and attention - Schneider, W. & Shiffrin, R. M. (1977). Psychological Review, 84(1), 1-66. [Reliability: High] ↩︎

  6. Dreyfus model of skill acquisition - Wikipedia / Dreyfus, S. E. & Dreyfus, H. L. (1980). [Reliability: High] (Original paper is peer-reviewed) ↩︎ ↩︎2 ↩︎3

  7. Cognitive Challenges in Human–Artificial Intelligence Collaboration: Investigating the Path Toward Productive Delegation - Fügener, A., Grahl, J., Gupta, A., & Ketter, W. (2022). Information Systems Research, 33(2), 678-696. [Reliability: High] ↩︎ ↩︎2 ↩︎3

  8. Tacit Knowledge - Wikipedia / Polanyi, M. (1958). Personal Knowledge. [Reliability: High] (Original is a classic philosophy book) ↩︎

  9. AI-Augmented Cognitive Task Analysis - MODSIM World (2025). [Reliability: Medium-High] ↩︎ ↩︎2

  10. Thinking, Fast and Slow - Kahneman, D. (2011). Farrar, Straus and Giroux. [Reliability: High] ↩︎

  11. Systems 1 and 2 thinking processes and cognitive reflection testing in medical students - Tay, S.W., Ryan, P., Ryan, C.A. (2016). Canadian Medical Education Journal. [Reliability: High] ↩︎

  12. Conditions for Intuitive Expertise: A Failure to Disagree - Kahneman, D. & Klein, G. (2009). American Psychologist, 64(6), 515-526. [Reliability: High] ↩︎ ↩︎2 ↩︎3 ↩︎4 ↩︎5

  13. Protecting Human Cognition in the Age of AI - Singh, A. et al. (2025). arXiv preprint. [Reliability: Medium-High] ↩︎ ↩︎2

  14. Is AI Creating Incompetent Experts? - IE Insights (2024). [Reliability: Medium] ↩︎ ↩︎2 ↩︎3

  15. Performance and Metacognition Disconnect when Reasoning in Human-AI Interaction - Fernandes, D. et al. (2024). arXiv preprint. [Reliability: Medium-High] ↩︎

AI・Technology
AI Expertise Metaknowledge Tacit-Knowledge Cognitive-Science Decision-Making
This post is licensed under CC BY 4.0 by the author.