Post
JA EN

The Polarization of 'Writes' and 'Write-Nots': Scientifically Verifying Paul Graham's Claims

Posted
By
15 min read
The Polarization of 'Writes' and 'Write-Nots': Scientifically Verifying Paul Graham's Claims
  • Target Audience: Software engineers, IT professionals interested in AI adoption
  • Prerequisites: Basic experience with AI tools such as ChatGPT, Claude, etc.
  • Reading Time: 25 minutes

Overview

In October 2024, Y Combinator founder Paul Graham published an essay titled “Writes and Write-Nots”1. His argument is clear and striking: with the advent of AI, society will polarize into “Writes” and “Write-Nots”—and this means a divide between “thinkers” and “non-thinkers.”

This article examines this claim based on research in cognitive science, educational psychology, and philosophy. In conclusion, Graham’s claims have substantial evidence, but several important assumptions and limitations also emerge.

Furthermore, we question the framework of “using AI as a tool” that Graham implicitly assumes. Drawing on philosopher Martin Buber’s dialogical philosophy (I-Thou vs I-It) and educational psychologist Vygotsky’s “Zone of Proximal Development” theory, we present the perspective that the quality of our relationship with AI determines growth.

Paul Graham’s Claims

Key Points of the Essay

Graham’s essay consists of three main claims1:

1. Writing is difficult—because thinking is difficult

“To write well you have to think clearly, and thinking clearly is hard.”

The difficulty of writing is not a matter of writing skill, but of clear thinking.

2. AI has made it possible to avoid writing

The act of “writing,” which previously could only be avoided through money (ghostwriters) or plagiarism, can now be easily avoided by anyone with AI. In schools and workplaces, people can delegate writing to AI.

3. Society will polarize into “Writes” and “Write-Nots”

Graham compares this to modern strength training. Physical strength was once necessary for daily life, but now only those who intentionally train possess it. Similarly, writing ability will become “a skill only those who choose it have.”

Leslie Lamport’s Quote

At the core of the essay, Graham quotes distributed systems authority Leslie Lamport:

“If you’re thinking without writing, you only think you’re thinking.”

This is the claim that writing is not merely recording thoughts, but thinking itself.

Scientific Verification: Supportive Evidence

Scientific evidence supporting Graham’s claims comes from multiple research areas.

1. Writing-to-Learn Research

“Writing-to-Learn” research has accumulated since the 1970s, with meta-analyses showing consistent results.

A meta-analysis of 56 experiments by Graham et al. (2020) showed that writing about content reliably promotes learning (effect size = 0.30)2. This effect:

  • Was equal across science, social studies, and math
  • Was equally effective from elementary to high school students
  • Was not moderated by writing activity, instruction, or evaluation characteristics

In other words, writing itself has a learning-promoting effect has been empirically demonstrated.

2. Extended Mind Theory

Philosophers Andy Clark and David Chalmers argued in their 1998 paper “The Extended Mind” that cognition does not occur solely within the brain3:

“Where does the mind stop and the rest of the world begin?”

Their answer is that cognitive processes “ain’t all in the head.” External objects that store information—calculations with pen and paper, diaries, computers—can function as part of the cognitive process.

This theory philosophically supports Lamport’s claim that writing is an act of “thinking” rather than “recording thoughts.”

3. Cognitive Offloading and AI

A 2025 study by Gerlich at SBS Swiss Business School found a significant negative correlation between frequent AI tool use and critical thinking ability decline4:

  • Younger people have higher AI tool dependence and lower critical thinking scores compared to older people
  • Cognitive offloading (delegating cognitive tasks to external tools) is the main cause of decline
  • However, higher education functions as a protective buffer against cognitive offloading

This research suggests that passive dependence on AI may affect cognitive abilities.

4. The Dual Nature of AI-Assisted Learning

A 2024 CHI Conference paper showed that generative AI use imposes metacognitive demands in three stages5:

flowchart TB
    subgraph A["Prompt Creation Stage"]
        direction TB
        A1["Self-awareness of task goals"]
        A2["Decomposition into subtasks"]
        A3["Explicit verbalization"]
    end

    subgraph B["Output Evaluation Stage"]
        direction TB
        B1["Validity judgment"]
        B2["Hallucination detection"]
        B3["Alternative consideration"]
    end

    subgraph C["Automation Strategy Stage"]
        direction TB
        C1["Prompt improvement"]
        C2["Task allocation judgment"]
        C3["Workflow integration"]
    end

    A --> B
    B --> C

Interestingly, this research shows that even “blindly delegating” to AI requires advanced cognitive work. That is, effective AI use presupposes exactly the “clear thinking” Graham mentions.

Scientific Verification: Critical Evidence

On the other hand, Graham’s claims have several important counterarguments and limitations.

1. Criticism of the “Great Divide” Theory Between Oral and Written Cultures

Linguists and anthropologists have objected to viewing oral and written cultures as binary oppositions.

Tannen (1988) rejected the interpretation that orality and literacy “form a dichotomous pair,” arguing that both are “complex and intertwined”6.

Anthropologist Ruth Finnegan states that “the once-confident claims about features said to distinguish literate from preliterate cultures are clearly wavering”6.

This criticism is important. The dichotomy of “Writes vs. Write-Nots” may oversimplify the complex continuity of cognitive abilities.

2. Questions About the Western Concept of “Critical Thinking”

The assumption that critical thinking is a universally valuable skill has itself been questioned7:

“What counts as critical thinking in the West—the techniques of analysis and evaluation, the style and linear structure of written argument—is in fact part of a Western cultural tradition.”

The cognitive ability to reason logically is universal to humans, but its valuation and expression vary by culture. Graham’s argument may implicitly rest on literate culture-centrism.

3. Criticism of Technological Determinism

Graham’s argument contains a “technological determinism” premise that technology unilaterally changes society. Critic Daniel Chandler points out that while technology may contribute to social change, it is not the only factor8.

From a social constructivist perspective, humans and social conditions shape the development and adoption of technology. How AI is used is determined not only by the intrinsic nature of technology, but by economics, regulation, and culture.

4. The Possibility of Appropriately Designed AI Support

Recent research suggests that AI support does not necessarily impair cognitive abilities9:

  • Students using integrated AI writing tools showed greater agency in the writing process and engaged in deeper knowledge transformation overall
  • Students adept at self-regulation use AI as a cognitive tool rather than a substitute
  • Those who use AI for deep dialogue and seeking explanations have enhanced learning, while those seeking direct answers have impaired learning

This shows that AI’s impact is not uniform, but depends on how it’s used.

Discussion: Beyond Technological Determinism

Is Polarization “Inevitable” or “A Choice”?

The core problem with Graham’s argument is depicting polarization as an inevitable consequence of technology. However, what research shows is that the choice of how to use it determines the outcome.

flowchart TB
    AI["AI Tools"]

    subgraph Active["Active Use"]
        direction LR
        A1["Invest thinking in prompt design"]
        A2["Critically evaluate output"]
        A3["Deepen understanding through dialogue"]
    end

    subgraph Passive["Passive Use"]
        direction LR
        P1["Seek direct answers"]
        P2["Accept output as-is"]
        P3["Skip thinking processes"]
    end

    AI --> Active
    AI --> Passive

    Active --> E1["Cognitive ability enhancement"]
    Passive --> E2["Cognitive ability decline risk"]

    style E1 stroke:#2ea44f,stroke-width:3px
    style E2 stroke:#d29922,stroke-width:3px

The Reality of the Digital Divide

Research on the AI divide shows a different axis of polarization than Graham10:

  • Access gap: Inequality in access to AI technology
  • Usage gap: Differences in skills, usage patterns, and participation
  • Outcome gap: Inequality in effectively utilizing AI outcomes

According to Randstad’s survey (2024, 12,429 people across 15 markets), 75% of companies have adopted AI, while only 35% of employees have received AI training in the past year11.

In this context, polarization is more likely to occur based on “whether one is in an environment to use AI effectively” rather than “whether to choose to write or not.”

The Buffer Effect of Higher Education

What’s interesting in Gerlich’s study is the finding that higher education functions as a “protective buffer” against the negative effects of cognitive offloading4.

This suggests that those who have already acquired critical thinking skills maintain their abilities even when using AI. In other words, the problem may not be AI’s existence itself, but education in basic cognitive skills.

Beyond “Tools”—Questioning the Quality of Relationship with AI

So far, we’ve discussed AI as a “cognitive tool.” However, philosopher Martin Buber’s perspective raises a fundamentally different question.

I-Thou vs I-It

Buber distinguished two modes of human relationship in his 1923 work I and Thou12:

I-It (Instrumental Relationship)I-Thou (Dialogical Relationship)
Treats the other as an object to useEncounters the other as a person
One-directional, manipulativeMutual, direct
Means for one’s own purposesThe relationship itself has value
“I” doesn’t changeBoth are transformed

According to Buber, in I-Thou relationships a new dimension of “between” emerges. There, the relationship produces something beyond the sum of individual contributions.

Interestingly, Buber’s dialogical philosophy has been applied to AI and robotics research in the journal AI & Society13. It has been proposed that the “ego-centric tradition” of Western philosophy tends to treat AI instrumentally, and a dialogical approach has been suggested.

Vygotsky’s “Zone of Proximal Development”—Can AI Become a “More Capable Other”?

Educational psychologist Lev Vygotsky argued that learning occurs through dialogical interaction. In his “Zone of Proximal Development” theory, collaboration with “more capable others” enables growth that cannot be achieved alone14.

The 2025 study “Dialogic Pedagogy for Large Language Models” applies this perspective to AI15:

“Framing LLMs not as authoritative answer-providers but as dialogical agents in education”

This research suggests that AI can function as a “co-constructor of knowledge.” However, important challenges are also noted:

“LLMs tend to give direct answers and are less likely to promote knowledge co-construction”

Relationship Quality Determines Learning Outcomes

Mentoring research shows that relationship duration and quality directly affect learning outcomes16:

  • Long-term mentoring relationships produce better academic, social, and psychological outcomes
  • AI is effective for “immediate customized feedback”
  • However, empathy, lived experience, and authentic connection exist only in human mentors

One researcher describes the essence of AI mentoring as “AI is the backbone, humans are the heart.”

Is a “Dialogical Relationship” with AI Possible?

flowchart LR
    subgraph Tool["I-It (Tool)"]
        direction LR
        T1["Efficient task completion"]
        T2["Information retrieval/generation"]
        T3["Cognitive load reduction"]
    end

    subgraph Partner["I-Thou (Dialogue)"]
        direction LR
        P1["Knowledge co-construction"]
        P2["Externalization and dialogue of thinking"]
        P3["Mutual questioning"]
    end

    Tool --> O1["Productivity improvement"]
    Partner --> O2["Cognitive ability transformation"]

    style O2 stroke:#2ea44f,stroke-width:3px

What research suggests is that whether we treat AI as “an object to use” or “a partner to dialogue with” may fundamentally change the quality of learning.

However, important limitations should also be recognized:

  1. Can AI truly become a “Thou”?
    • Buber’s I-Thou presupposes mutual transformation
    • AI doesn’t “transform” (at least currently)
    • Some researchers are skeptical, but there is research exploring the possibility of “pseudo-I-Thou”
  2. Risk of Anthropomorphization
    • Treating AI as a person may create emotional dependence17
    • Risk of decreased critical thinking
    • Impact on interpersonal communication skills
  3. Importance of Design
    • “Dialogical AI tutors” need to be designed to gradually reduce support rather than give direct answers
    • An approach of “Now try it yourself. I’ll help if you get stuck”

Implications for Engineers

When using GitHub Copilot or Claude Code, the awareness of “using a tool” versus “having a dialogue” may produce different gains.

I-It approach (Tool):

1

"Write this code" → Copy-paste output → Move to next task

I-Thou approach (Dialogue):

1

"What do you think of this design?" → Discussion → "Why that choice?" → Deep dive → "Other approaches?" → Comparative examination → Make your own judgment

With the latter approach, one’s own thinking transforms through “dialogue” with AI. This is close to what Buber calls the experience of “between.”

Summary

Paul Graham’s “Writes and Write-Nots” raises an important issue about the relationship between AI and cognitive abilities. Based on scientific verification, we can say the following:

Points Where Graham’s Claims Are Supported

  1. The close relationship between writing and thinking: Writing-to-Learn research and Extended Mind theory support that writing is not merely a record of thought, but the thinking process itself
  2. Risk of cognitive offloading: Multiple studies have shown that passive dependence on AI may affect critical thinking abilities

Limitations of Graham’s Claims

  1. Oversimplification of binary opposition: The classification of “Writes vs. Write-Nots” ignores the continuity and diversity of cognitive abilities
  2. Technological determinism premise: AI’s impact depends not only on technology’s intrinsic nature, but on usage, education, and social context
  3. Cultural bias: Arguments equating literate culture with logical thinking contain Western-centric premises
  4. Underestimates the possibility of active use: Appropriately designed and used AI support can enhance cognitive abilities

Perspectives Graham Overlooks

  1. The limitation of the “tool” framework itself: Whether we treat AI as a tool (I-It) or approach it as a dialogical partner (I-Thou oriented) may fundamentally change the quality of learning gained

Buber’s dialogical philosophy, Vygotsky’s Zone of Proximal Development theory, and mentoring research suggest that relationship quality determines growth. Through “dialogue” with AI, the human side is transformed—this possibility cannot be captured within the “tool use” framework.

Implications for Engineers

While accepting Graham’s warning, it’s important not to fall into technological determinism. When using AI coding tools:

  • Intentionally preserve thinking processes: Think about design yourself before having code generated
  • Critically evaluate output: Don’t use generated code as-is; understand it before adopting
  • Deepen understanding through dialogue: Ask AI “why this implementation?” and consider alternatives
  • Try treating AI as a “dialogue partner”: Start with “what do you think about this?” rather than “do this”

Ultimately, if polarization occurs, it won’t be between “Writes and Write-Nots” or “AI users and non-users,” but between “those who transform themselves through AI and those who surrender their transformation to AI.”

And that divergence point may lie in our attitude of whether we use AI as “It” or face it as “Thou.”

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. Writes and Write-Nots - Paul Graham (2024). [Reliability: Medium] ↩︎ ↩︎2

  2. The Effects of Writing on Learning in Science, Social Studies, and Mathematics: A Meta-Analysis - Graham, S., Kiuhara, S. A., & MacKay, M. (2020). Review of Educational Research. [Reliability: High] ↩︎

  3. The Extended Mind - Clark, A. & Chalmers, D. (1998). Analysis, 58(1), 7-19. [Reliability: High] ↩︎

  4. AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking - Gerlich, M. (2025). Societies, 15(1), 6. [Reliability: High] ↩︎ ↩︎2

  5. The Metacognitive Demands and Opportunities of Generative AI - Tankelevitch, L. et al. (2024). CHI Conference on Human Factors in Computing Systems. [Reliability: High] ↩︎

  6. Oralities & Literacies – Chapter 1 – Synthesizing the Orality Debate - International Orality Network. [Reliability: Medium-High] ↩︎ ↩︎2

  7. The role of critical thinking in academic writing - ERIC Database. [Reliability: Medium-High] ↩︎

  8. A Critical Analysis of Technological Determinism Theory - Preprints.org. [Reliability: Medium] ↩︎

  9. Human-AI collaboration patterns in AI-assisted academic writing - Taylor & Francis (2024). Studies in Higher Education. [Reliability: High] ↩︎

  10. Mind the AI Divide: Shaping a Global Perspective on the Future of Work - United Nations (2024). [Reliability: High] ↩︎

  11. AI skills gap widens - Randstad (2024). Understanding Talent Scarcity: AI and Equity Report. [Reliability: High] ↩︎

  12. I and Thou - Buber, M. (1923/1958). Scribner. [Reliability: High] ↩︎

  13. Rethinking the I-You relation through dialogical philosophy in the Ethics of AI and robotics - AI & Society (2017). [Reliability: High] ↩︎

  14. Vygotsky’s Zone of Proximal Development: Instructional Implications - ERIC Database. [Reliability: Medium-High] ↩︎

  15. Dialogic Pedagogy for Large Language Models: Aligning Conversational AI with Proven Theories of Learning - Beale, R. (2025). arXiv preprint. [Reliability: Medium-High] ↩︎

  16. Artificial intelligence as a mentor in the graduate online classroom - Emerald Publishing, AI in Education (2024). [Reliability: High] ↩︎

  17. Anthropomorphic response: Understanding interactions between humans and artificial intelligence agents - Computers in Human Behavior (2023). [Reliability: High] ↩︎

AI・Technology
AI Writing Cognition Critical-Thinking Paul-Graham
This post is licensed under CC BY 4.0 by the author.