Post
JA EN

Feeling You Must Relearn Everything for the AI Era? — Three Practices to Keep Your Thinking Agency

Posted
By
20 min read
Feeling You Must Relearn Everything for the AI Era? — Three Practices to Keep Your Thinking Agency
  • Target audience: Knowledge workers and engineers who feel they must start over and learn from scratch because of AI advances
  • Prerequisites: Some experience using AI tools like ChatGPT in your work
  • Reading time: 20 minutes

Overview

“In the AI era, the old ways no longer work. I need to start from scratch and acquire a new way of thinking.” If something tightens in your chest every time you see that kind of message on social media or at meetups, you are not alone. A Pew Research Center survey found that 52% of U.S. workers are worried about how AI will affect the workplace, and 32% expect fewer job opportunities for themselves in the long run1. The anxiety is not a fantasy — it is a data-backed collective phenomenon.

But the conclusion that anxiety tends to push us toward — “set aside everything I have built and hand my thinking process over to a new AI-centered way of working” — is likely to backfire, according to multiple studies. Higher AI usage correlates with lower critical-thinking scores, and once AI assistance is removed, expert judgment accuracy drops — a “deskilling” effect documented in several research streams.

The answer of this article is simple. It is fine to acknowledge the anxiety. There are areas where relearning genuinely matters. But the discriminating eye — the ability to decide what to keep and what to replace, which is your thinking agency — must not be handed over. This article first lays out six skills that actually become more valuable in the AI era (the ones we are tempted to discard out of panic), then offers three practices for keeping agency, and ends with the “85% known + 15% novel” optimal-learning ratio from Wilson et al. (2019). For the deeper research, follow the internal links to existing posts on this blog.

Where does the “I have to start over” anxiety come from?

The collective anxiety is real

Pew Research Center surveyed 5,273 employed U.S. adults from October 7–13, 2024. The result: 52% expressed concern about AI’s future impact in the workplace, 32% expected fewer long-term job opportunities, and 33% reported feeling “overwhelmed” by AI1.

In other words, AI anxiety is not a sign of mental fragility — it is a collective psychological state shared by more than half of working people. The messages flowing through your timeline — “the old ways are dead,” “I was thinking on autopilot, now I am rebuilding from zero” — are merely surface expressions of this same anxiety.

The moment anxiety becomes overreaction

The problem is not the anxiety itself. It is the mistranslation that happens when anxiety gets converted into action:

  • “Change is fast” → “Past experience is outdated and unusable”
  • “AI can do more things” → “I must hand thinking process to AI and reorganize around it”
  • “Unlearning is important” → “Reset everything I have built and install a new way”

Each of these translations involves overreaction. Fast change does not mean past knowledge becomes invalid. Expanded AI capability does not mean human judgment becomes unnecessary. Unlearning is not about resetting — it is the concept of letting go selectively (covered in detail in Unlearning and Relearning in the AI Era).

A typical form of overreaction is: “The old way of thinking won’t work, so I will hand the reins of thought to AI and reposition myself as someone who uses AI well.” It sounds reasonable, but it is the entry point to handing over thinking agency.

Why “swap out your thinking entirely” backfires

The reasons why anxiety-driven “wholesale replacement of thinking” backfires are explained by multiple research streams. This article gives only the conclusions — follow the internal links for the underlying mechanisms.

1. Cognitive offloading erodes critical thinking

A 2025 study by Gerlich in Societies (MDPI), based on 666 participants, found a strong negative correlation (r = -0.68) between AI usage frequency and critical-thinking scores2. The more people use AI, the more their thinking gets externally offloaded — and the critical-thinking muscle weakens.

That said, this correlation does not establish causation. It is also plausible that people with weaker critical thinking are more inclined to depend on AI. For a fuller discussion, see Cognitive Offloading and Critical Thinking.

2. Performance collapses the moment AI is removed

A Lancet medical-AI study reported that after AI-assisted endoscopy was introduced, physicians’ lesion-detection rate on non-AI cases dropped from 28.4% to 22.4%3. The crucial finding is not the AI-assisted performance — it is the performance without AI that fell.

This “deskilling the moment AI is gone” pattern is not abstract for engineers. The day AI goes down, the day a code review suddenly needs a human eye, your raw capability is exposed. For details, see The AI Deskilling Paradox.

3. AI’s benefit reaches only those with high metacognition

A field experiment with 250 employees at a Chinese technology consulting firm found that only employees with high metacognition (the ability to plan, monitor, and revise their approach) saw creativity gains from ChatGPT use4. The optimistic narrative — “AI democratizes creativity for everyone” — is contradicted by the research.

In other words, fully outsourcing your thinking to AI does not work without metacognition as a foundation. Discard the foundation and keep only the AI, and creativity does not rise — you just become a passive worker. For details, see The Metacognition–AI Creativity Gap.

flowchart TB
    A["Anxiety: 'Must relearn<br>from scratch'"]
    B["Outsource thinking<br>entirely to AI"]
    C1["Cognitive offloading<br>erodes critical thinking"]
    C2["Raw skill exposed<br>when AI is removed"]
    C3["Metacognition foundation<br>never develops"]
    D["You fall behind<br>even more"]

    A --> B
    B --> C1
    B --> C2
    B --> C3
    C1 --> D
    C2 --> D
    C3 --> D

Accept the anxiety — a psychological preprocessing

Before getting to the practices, one preliminary note. You don’t need to suppress the anxiety.

Anxiety is in fact a sign that your change-detection sensor is working. Given that more than half of all workers share the same feeling, it is enough to acknowledge “I am not the only one who feels left behind” and start there.

The problem is not the anxiety itself — it is the conversion of anxiety into the extreme action of “replacing your entire way of thinking”. If you can separate “feeling the anxiety” from “deciding the action,” the practices that follow become easy to implement.

Psychological approaches to AI anxiety itself (growth mindset, self-compassion, etc.) are covered in Engineer’s Learning Anxiety — Three Pillars to Overcome It. This article goes one step further, using cognitive-science evidence and concrete practices to examine how to keep your thinking agency.

Skills that become more valuable in the AI era

“My past experience won’t work anymore” is often a factual misperception. The skills below are easy to set aside in a panic, yet they tend to gain value, not lose it, in the AI era. Six representative ones:

1. The ability to structure problems

If you can’t decide what to ask the AI, you won’t get useful output from it. Decomposing big problems, prioritizing, and organizing the points becomes a scarce resource in the AI era — discussed in detail in The Real Identity of “Generalists” Is the Power to Ask Questions.

2. Domain knowledge and a verifying eye

Only those with deep domain knowledge can spot when AI output is “plausible but wrong.” Because AI usage is the default, the value of domain knowledge — the basis for verifying its output — goes up. The misconception “if AI does it, verification can be light” was rebutted in The More You Delegate to AI, the Heavier Verification Becomes.

3. Judgment criteria and values

People who can answer “what should we build?” and “why are we building this?” are not going away. Your values, the aesthetic sensibility built through your work — these are domains AI can’t take over. This pairs with Practice 2 (“Keep the why in your own hands”) later.

4. Interpersonal skill and accumulated trust

Customer relationships, trust with colleagues, the ability to move a team — these are not just irreplaceable by AI; they become the central asset of who you can mobilize in an AI-augmented organization. Panicking and over-investing solely in “personal AI fluency” risks eroding this long-term asset.

5. The experience of having “done it with your own hands”

Pattern recognition isn’t built from abstract knowledge — it forms through accumulated concrete failures and successes (see How Engineers’ Pattern-Recognition Asset Transfers to the AI Era). The moment you decide “AI writes it, so I don’t have to,” that accumulation stops. The fact that experts are precisely the ones who get the most from AI is also discussed in The Value of Experience in the AI Era.

6. The ability to write and articulate

The skill of writing precise prompts for AI is, in the end, an extension of the ability to articulate your own thinking. It isn’t that “writing is no longer needed” — it’s that “what you write to has changed.” If anything, those weaker at articulation tend to fall into a loop: vague instructions to AI, vague outputs back.

These are not “things to relearn” — they are assets you can carry directly into the AI era. The moment you pick wholesale replacement out of panic, these get pushed aside. To say it another way: practice in the AI era is not “acquire something new from zero,” but “lay your AI-usage approach on top of the assets you carry over.” The three practices in the next section are all about that “laying on top.”

Three practices to keep your thinking agency

Practice 1: Use a “thinking sprint” to put your own hypothesis first

Study Hacker (February 2026) proposes a concrete procedure called the “thinking sprint” for keeping your thinking faculty alive alongside AI5:

  1. First 5 minutes: Without AI, think on your own. Sketch a structure, jot down a tentative answer.
  2. Next 4 minutes: Ask AI: “What’s missing?” “What’s an alternative?” “Help me organize the points.”
  3. After that: Examine the AI output and decide yourself what to adopt and what to discard.

The order matters. Many people start by asking AI immediately, then think from AI’s answer as the starting point. That is a textbook case of cognitive offloading — without your own hypothesis, you simply assimilate to the AI’s output.

Putting your own hypothesis first has three effects:

  • You can compare AI output to your own view (assimilation becomes dialogue)
  • You become better at noticing what AI missed
  • Your thinking muscle keeps getting used, blocking deskilling

The timing does not have to be strict. What matters is keeping the order “me first, AI later.”

Practice 2: Keep the “why” to yourself — don’t hand it to AI

Let AI handle the “how,” but never the “why” — this is the most important dividing line for keeping thinking agency.

TypeSafe to delegateDo not delegate
“How”Implementation steps, syntax, routine tasks—
“What”Option generation (have AI produce multiple alternatives)Final selection
“Why”—Purpose, priorities, value judgments

“Why this design?” “Why this feature?” “Why in this order?” — the moment you can no longer answer these yourself, you have lost agency over your work. Asking AI “why?” is itself a useful hypothesis-generating move, but if you accept the AI’s explanation without holding your own judgment in reserve, you are simply overwriting your own judgment with someone else’s. Even after asking AI, you need to compare it against your own “why.”

In practice, adding a one-line “why” of your own to your prompts works well. The form: “For X purpose, I want Y.” Just that, and you can immediately tell, when AI output comes back, whether it serves your purpose.

Practice 3: Develop a discriminating eye

The last practice is the right operation of unlearning. Switch from “relearn from scratch” to “selecting what to replace and what to keep.”

Selection guide:

  • Keep: The ability to see problem structure, judgment criteria, domain knowledge, accumulated human relationships and trust, the values you cherish (these largely map to the six skills above)
  • Replace through relearning: Specific tool-level operating details, knowledge that depended on rote memorization, the norm “I should do everything myself,” attachment to past success patterns

When framing the selection question, ask “is this necessary for my judgment, or not?” instead of “can AI replace this, or not?” The former question shifts year by year as tools evolve; the latter is anchored in the essence of your work, so it stays stable.

The act of selection is itself a metacognitive exercise. By going through the process of deciding what to keep, you become able to articulate the core of your work. People who are caught up in “I must relearn from scratch” often haven’t found the time to articulate their own core.

A sustainable pace — “85% known + 15% novel” is optimal

After separating “to relearn” from “to keep,” the next decision is at what ratio you bring in the new. There is, in fact, cognitive-science evidence that answers this directly.

Wilson’s “85% rule” — the sweet spot is a 15% error rate

A 2019 paper by Wilson, Shenhav, Straccia, and Cohen in Nature Communications mathematically derived the optimal difficulty for learning6. For a broad class of stochastic-gradient-descent-based learning algorithms, the optimal error rate is about 15% (accuracy ≈ 85%). Too easy and learning stalls; too hard and frustration kicks in — there is a sweet spot in between.

StateRatio knownRatio novelResult
Too easy100%0%Learning stagnates
Optimal85%15%Learning maximized
Too hard≤50%≥50%Cognitive saturation, dropout

The “85% rule” was demonstrated for both artificial neural networks and biologically plausible neural networks thought to describe animal learning, and applies especially well to perceptual learning and pattern-recognition-type learning6.

Why this matches the article’s theme exactly

Wilson’s rule meshes precisely with this article’s argument:

  • Wholesale replacement is the worst case: 100% novel is extremely inefficient for learning. Anxiety-driven “relearn from scratch” is also inefficient in cognitive-science terms.
  • Zero is also the worst case: 100% known (= replacing nothing) also stops learning.
  • Accumulated knowledge is the foundation for learning: It is precisely because 85% is known that the remaining 15% can be learned meaningfully.

Translated to engineering practice:

  • Work in a familiar tech stack, with about 15% of that work going to a new technology or new way of thinking
  • Continue using existing languages while weaving in 15% adjacent domains or new libraries
  • When introducing AI tools, replace about 15% of an existing workflow at a time — this keeps deskilling in check and maximizes learning efficiency

“85% comes from accumulated assets” connects directly to the article’s core argument — build on the six carry-over skills from the previous section as your foundation.

Other ratio conventions as supporting lines

On the time-allocation axis, The Career Strategy to Start Before Cognitive Decline introduces the “10% rule” — allocating 10% of work time to learning (4 hours per 40-hour week). This is a different axis from Wilson’s rule (content ratio vs. time ratio), and they coexist: use 10% of your work time for learning, and within that learning, combine 85% known with 15% novel.

Other conventions, for reference:

ConventionAxisContentEvidence strength
Wilson’s 85% rule6Content ratio85% known + 15% novel is the sweet spotModerate–high (peer-reviewed in Nature Communications, but derived from binary-classification tasks — full transfer to complex human learning needs caution)
10% time ruleTime ratioAllocate 10% of work time to new learningWeak (industry convention, no academic source)
70-20-10 model7Activity ratio70% experience, 20% from others, 10% formalWeak (1996 publication based on retrospective self-report; subject to critique8)
Google’s “20% time”9Time ratio20% of work time on self-chosen projectsFamous as a policy, no academic validation
Japan IT industry self-study survey10Actual practice81.4% engage in self-study; majority do 1–5 hours per weekUseful as practice data, doesn’t show an optimum

Ratios on different axes can coexist. For example: “allocate 10% of work time to learning (time axis), and within that learning combine 85% known with 15% novel (content axis).”

Putting these conventions together, “keep 80–90% as your foundation and feed in 10–20% of new territory continuously” is the rough guideline that both strong scientific evidence and multiple industry conventions support.

The exact number isn’t sacred. 85/15 or 90/10 — what matters is “don’t make it zero / don’t make it total / keep it going.” Sustaining that rhythm beats any panic-driven wholesale replacement.

Summary

The anxiety of “I have to relearn everything for the AI era” is a real psychological phenomenon shared by more than half of working people. But the moment that anxiety converts into “replacing my entire way of thinking,” the traps that research points to — cognitive offloading, deskilling, the metacognition gap — catch your foot.

What you need to keep agency is not a wholesale switch in your way of thinking, but the discriminating eye to decide what to keep and what to replace. Put together, this article’s argument:

  • You have more carry-over assets than you think: problem structuring, domain knowledge, judgment criteria, interpersonal skill, hands-on experience, articulation — these six gain value in the AI era
  • Three practices to keep agency: (1) thinking sprint to put your own hypothesis first, (2) keep the “why” in your own hands, (3) ask “is this necessary for my judgment?” as the selection criterion
  • Cognitive-science basis for a sustainable pace: Wilson’s 85% rule (85% known + 15% novel) is the learning sweet spot. “Build on existing accumulation as foundation and keep feeding in 15% of new content” is optimal

You don’t need to deny the anxiety. Acknowledge it, and still keep deciding your actions calmly — once you can separate those two, you won’t fall behind in the AI era. The ones who fall behind are not those who feel anxious, but those who, driven by anxiety, hand over their own judgment.

For more on related themes:

References

References cited by footnote number in the article body.

Other references (not cited by number)

  1. U.S. Workers Are More Worried Than Hopeful About Future AI Use in the Workplace — Pew Research Center (published Feb 25, 2025). Based on a survey of 5,273 employed U.S. adults conducted October 7–13, 2024. [Reliability: High] ↩︎ ↩︎2

  2. AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking — Gerlich, M. (2025). Societies, 15(1), 6. MDPI. Mixed-methods study with 666 participants. Correlation between AI usage frequency and critical-thinking score: r = -0.68 (p < 0.001). Causality cannot be confirmed because the study is cross-sectional. [Reliability: Moderate–High] ↩︎

  3. Wang, P. et al. (2025). Medical AI study published in The Lancet. After continued use of AI-assisted endoscopy, physicians’ lesion-detection rate on non-AI cases dropped from 28.4% to 22.4%. For detailed data and source, see The AI Deskilling Paradox. [Reliability: High] ↩︎

  4. Lu, J., Sun, S., Li, Y., Foo, M.-D., & Zhou, J. (2026). Field experiment with 250 employees at a Chinese technology-consulting firm. Creativity gains from ChatGPT use occurred only in employees with high metacognition. For detailed data and source, see The Metacognition–AI Creativity Gap. [Reliability: Moderate–High] ↩︎

  5. AI時代の「思考筋トレ」──考える力を失わないためのコツ — Marika Hashimoto, STUDY HACKER (February 6, 2026). Proposes the “thinking sprint” (5 min self → 4 min AI → self-judgment) as a concrete procedure. (Japanese article.) [Reliability: Moderate] ↩︎

  6. The Eighty Five Percent Rule for optimal learning — Wilson, R. C., Shenhav, A., Straccia, M., & Cohen, J. D. (2019). Nature Communications, 10, 4646. Mathematically derives that the optimal error rate is approximately 15% (accuracy ≈ 85%) for a broad class of stochastic-gradient-descent-based learning algorithms; demonstrated for both artificial neural networks and biologically plausible neural networks. [Reliability: Moderate–High (peer-reviewed paper, but derived from binary-classification tasks; full transfer to complex human learning needs caution)] ↩︎ ↩︎2 ↩︎3

  7. 70/20/10 model (learning and development) — Wikipedia. A learning-allocation model originating from a retrospective self-report study of about 200 executives by McCall, Lombardo, and Eichinger at the Center for Creative Leadership; the 70/20/10 ratio was publicly formalized in Lombardo & Eichinger, The Career Architect Development Planner (1996). [Reliability: Moderate — widely referenced industry model, but the underlying methodology is weak] ↩︎

  8. 70:20:10: Where Is the Evidence? — Association for Talent Development (ATD) blog. A critique pointing out that the 70-20-10 model is “neither a scientific fact nor a recipe for development,” and that its self-report basis carries methodological limitations (honesty, introspection, sampling bias). [Reliability: Moderate–High] ↩︎

  9. Google’s “20% Time” Policy: Case Study — Naresh Sekar, Medium. Covers Google’s 20% time policy (introduced in the early 2000s, with 3M’s 15% time as an earlier example), the origin of Gmail / Google News / AdSense, and the later shift toward structured innovation weeks. [Reliability: Moderate] ↩︎

  10. 自己研鑽を行っているIT人材は約80%以上！年代・年収ごとに見る自己研鑽の実態 — Geekly Media (2022). Survey showing 81.4% of Japanese IT professionals engage in self-study, with a majority dedicating 1–5 hours per week. (Japanese article.) [Reliability: Moderate] ↩︎

AI・Technology
AI Unlearning Deskilling CriticalThinking Career
This post is licensed under CC BY 4.0 by the author.