Why healthy doubt beats AI confidence theater 🔗 **Source:** [Why healthy doubt beats AI confidence theater](https://joost.blog/healthy-doubt/) > [!note]+ Document Notes > tags: [AI, education, workforce] > author: Joost de Valk > source: joost.blog > reference: [https://joost.blog/healthy-doubt/] > date created: 2026-02-23 14:00 > > --- > > # 🏷️ Title: Why healthy doubt beats AI confidence theater > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Include links cited in or relevant to the article. --- ## 📝 Summary - What happened. - The author argues that modern AI systems project extreme confidence while lacking self-awareness of ignorance, and that this reshapes work, education, and what “junior” means. - Why it matters. - “Average” output is now cheap and abundant; the scarce skill is human judgment, decomposition, and the courage to doubt AI’s polished answers. - Key stakeholders. - Students, educators, junior and senior workers (especially developers), managers/founders, and institutions designing curricula or hiring practices. - Risks and tradeoffs. - Over-reliance on AI erodes pathways to senior expertise, rewards shallow “prompting” over deep thinking, and encourages outsourcing responsibility to machines. - Bottom line. - AI makes average workers obsolete and excellent workers irreplaceable; cultivating skepticism, philosophical thinking, and “pilot not passenger” skills is now essential. --- ## 📖 Main Content - The author contrasts AI’s “confidence theater” with philosophical traditions of doubt: Socrates’ “I know that I know nothing,” Arendt’s distinction between thinking and knowing, and Kant’s “Sapere aude.” - AI is framed as a “knowing machine” that can generate polished answers but cannot tell when it’s wrong; hallucinations and truths are delivered with the same poise. - Because AI can now produce passable essays, code, and plans instantly, “average output has no market value anymore.” Human value shifts from production to evaluation and judgment. - The “junior paradox”: historically, people became senior by doing large amounts of junior work. If AI does that work, future seniors have no training ground, yet entry-level expectations are rising (“you can do what AI does plus catch where it fails”). - The “pilot vs passenger” model: effective professionals use AI like autopilot—understanding and supervising it—rather than riding as passengers who are helpless when it fails. - The “Descartes reflex” is introduced as the habit of decomposing and verifying complex outputs, akin to debugging or parsing Latin. Classical education is reframed as training for AI-era cognitive skills. - Kant’s “Sapere aude” is updated: instead of church and king, we must stop outsourcing thinking to “algorithm and model.” You may outsource work to AI, but not responsibility. - Concrete recommendations follow for educators (grade process, not product), managers (hire for judgment), entrants to the workforce (embrace “I don’t know yet”), and everyone (treat philosophy as cognitive self-defense). - The essay closes with the “uncomfortable truth”: AI commoditizes average workers but amplifies the value of those with deep, hard-won understanding. --- ### 🗒️ Notes - “AI will confidently sign off on anything. The question is whether you will too.” — sets up the central tension between AI’s confidence and human responsibility. - The author’s context is advising a Dutch gymnasium’s AI council, tying AI’s impact on the workforce back to classical education in Latin, Greek, and philosophy. - “Ask Claude, Gemini, or ChatGPT anything and you’ll get a polished, articulate, supremely confident answer. Even when it’s wrong.” — highlights the confidence problem. - Arendt’s thinking vs knowing: AI excels at producing answers but cannot “examine those answers, sit with uncertainty, [or] refuse to stop at the surface.” - “We’ve built machines that perform certainty, and we’re raising a generation that’s learning to mistake that performance for truth.” — core societal risk. - “Average output has no market value anymore.” Passable essays, working code, basic analysis: all now cheap commodities generated by AI. - Human edge: “It can’t feel that something is off. It can’t exercise judgment born from years of getting things wrong and learning why.” - Junior paradox: seniority historically requires years of “junior work” (bad code, many parsed sentences). AI now erodes that ladder by doing junior work itself. - “Entry-level now means you can do what AI does plus catch where it fails.” — a new, higher bar for juniors. - Pilot vs passenger: real pilots understand and oversee autopilot; good workers must supervise AI like a powerful but fallible intern. - “The market is firing passengers and hiring pilots… eliminating executors and desperately seeking directors.” — preference for architects over pure executors. - Descartes reflex: break problems into small pieces, verify each, then rebuild — the debugging mindset AI threatens by making outputs “look” right. - Classical education (e.g., parsing Latin) is reframed as practice in decomposition and reconstruction—exactly what’s needed to interrogate AI output. - Kant’s Sapere aude: stop outsourcing thinking to authority; today’s authority is “algorithm and model,” not church and king. - “You can use AI to outsource the work. You can’t outsource the responsibility.” — ethical boundary of AI use. - Educator advice: stop grading only the final product; interrogate the process with questions like “Why this word and not that one?” and “How do you know this is right?” - Manager advice: “Hire for judgment, not output volume… The person who can tell you why the AI’s answer is wrong is worth ten people who can generate AI answers quickly.” - Worker advice: your value lies “in the gap between ‘this looks right’ and ‘this is right’” and the willingness to say “I don’t know yet.” - Philosophy becomes “cognitive self-defense”: deconstructing arguments, questioning assumptions, and judging under uncertainty. - Final claim: “AI makes the average worker obsolete, but the excellent worker irreplaceable… The wind is in the sails. Humans still set the course.” --- ## 📚 References - Christina Aguilera, “AI Confidence Theater” (LinkedIn leadership piece) - Immanuel Kant, “Answering the Question: What Is Enlightenment?” (1784) - Hannah Arendt, works distinguishing thinking vs knowing (esp. “The Life of the Mind”) - Aristotle, “Nicomachean Ethics” (practice, virtue, habituation) - René Descartes, “Discourse on Method” (decomposition of problems) - Joost de Valk, “The death of Code Copyright and the rise of the Architect” (referenced prior post) - Original article: “Why healthy doubt beats AI confidence theater” — https://joost.blog/healthy-doubt/ --- ## 🗣️ Paraphrase - Paragraph A paraphrase. - Modern AI systems answer any question with polished confidence, regardless of correctness, and society risks mistaking that performance for genuine understanding. - Paragraph B paraphrase. - As AI automates mediocre work, human value shifts toward critical thinking, skepticism, and the ability to dissect and correct AI output—skills long cultivated by philosophy and classical education. --- ## 📰 Newsletter AI tools can now produce “good enough” content on command—but Joost de Valk argues that’s exactly why healthy doubt matters more than ever. In his latest essay, he connects classical philosophy to today’s AI-powered workplace, showing why WordPress professionals must be pilots, not passengers, when using tools like GitHub Copilot or content generators. Worth a read if you’re rethinking workflows, training juniors, or designing AI-assisted editorial processes in WordPress. --- ## ✅ Action Items #### DevRel - Prioritize talks and demos that show debugging and verifying AI output, not just generating it. - Create workshops teaching “Descartes reflex” skills: code review, reasoning checks, and prompt-plus-judgment patterns. - Produce example repos and quickstarts that pair AI-assisted code with human review checklists. - Run campaigns contrasting “average AI output” vs curated, explained solutions to highlight human value. - Facilitate community office hours focused on reading and critiquing AI-generated patches. #### OSPO - Draft guidance clarifying that AI-generated contributions still require human accountability and review. - Update contribution policies to document provenance and review of AI-assisted code. - Integrate security scanning and license checks that account for AI-suggested dependencies. - Maintain inventories (SBOMs) that note where AI-assisted changes enter critical components. - Educate contributors and maintainers on risks of blindly accepting AI patches and the need for verification. #### Community - Offer “architecture clinics” where maintainers walk through how they evaluate AI-generated proposals. - Design contributor onboarding that teaches how to interrogate AI output instead of pasting it verbatim. - Encourage sponsors to fund mentorship for juniors learning real debugging and design, not just prompting. - Coordinate cross-project incident drills involving AI-induced bugs and their remediation. #### WordPress - Core: Emphasize code review standards for AI-assisted patches and require clear rationale in tickets. - Core Performance: Benchmark AI-generated optimizations and validate with profiling before merging. - Core AI: Establish responsible-use guidelines, evaluation sets, and failure-mode exemplars. - Core Program Team: Add regular reviews of how AI affects contributor pathways (especially juniors). - Mobile: Document where AI helps with boilerplate while ensuring humans own UX and edge cases. - Playground: Provide sandboxes for experimenting with AI-generated code in a safe environment. - Community: Include “AI skepticism” modules in meetup kits and WordCamp talks. - Training: Build lesson plans on reading, testing, and refactoring AI-generated code/content. - Marketing: Position WordPress AI features as copilots that still require human editors. - Support: Capture common AI-related support issues and encode them into docs and training. - Design: Audit AI-generated UI ideas with human-led usability testing. - Documentation: Add examples explaining how to verify AI-written snippets and queries. - Photos: Clarify policies around AI-generated media and attribution. - TV: Encourage talks on AI literacy, doubt, and responsible usage. - Polyglots: Evaluate AI-assisted translations with human reviewers, documenting known failure modes. - Accessibility: Ensure AI-generated interfaces and content are checked against WCAG. - Meta: Provide tooling to flag suspiciously generic (likely AI) contributions for extra review. - CLI: Offer safe scaffolds but require explicit human review steps for generated commands. - Hosting: Share best practices for AI-assisted configuration, logging, and debugging. - Tide: Consider how AI-generated plugins/themes affect scoring and verification. - Openverse: Clarify inclusion rules and labeling for AI-generated media. - Themes: Mentor authors on using AI for drafts while maintaining human-led design systems. - Plugins: Strengthen security reviews where AI may have introduced subtle flaws. #### The WPOCC - Programs: Fund fellowships teaching critical AI use and architectural thinking to maintainers. - Projects: Run documentation sprints on AI best practices and governance. - Funding experiments: Pilot grants for tools that make AI output more inspectable and auditable. ### FAIR - Encourage local evaluation and metadata to track AI involvement in code and content. - Publish governance artifacts outlining acceptable AI use and transparent funding of AI tooling. --- ## 🧪 Hypotheses - H1: Workers trained in decomposition and philosophical reasoning will detect AI errors more reliably than those trained only in tool usage. - H2: Organizations that hire for judgment over output volume will experience fewer AI-induced production incidents. - H3: Curricula that assess students on process, not just product, will better prepare them for AI-saturated workplaces. - H4: Over-reliance on AI for junior tasks will reduce the future supply of true senior experts. - H5: Explicitly teaching “pilot vs passenger” mental models will improve responsible AI usage. --- ### 📊 Methodology - Run controlled experiments comparing groups with and without explicit training in the “Descartes reflex” on tasks involving flawed AI outputs. - Conduct longitudinal studies across companies adopting AI, tracking incident rates vs hiring practices (judgment-focus vs volume-focus). - Compare outcomes for students in programs with process-based grading vs traditional essay grading in AI-rich environments. - Analyze career trajectories in teams heavily automating junior tasks vs those preserving manual practice opportunities. - Use surveys and behavioral assessments to measure “pilot vs passenger” attitudes and correlate them with real-world error detection. --- ### ❓ Questions - How can we design scalable training pathways that preserve junior learning while leveraging AI? - What signals reliably indicate human judgment in hiring beyond portfolio volume? - How should educational accreditation adapt in a world where AI can write any assignment? - Which parts of software and content pipelines should remain strictly human-signed? - How can open-source communities transparently track AI involvement without deterring contributors? --- ### 🔮 Future Research - Develop metrics for “judgment density” in roles and organizations. - Explore tools that make AI reasoning chains visible and critique-able. - Investigate hybrid human–AI apprenticeship models that maintain deep practice. - Study long-term cognitive impacts of frequent AI reliance on problem-solving. - Examine cross-cultural differences in trust of AI authority and how they affect adoption. --- ### 💡 Implications - Practice: Professionals must cultivate debugging, decomposition, and argument analysis as core skills alongside tool fluency. - Policy: Institutions may need guidelines assigning final responsibility for AI-assisted decisions. - Security: Attackers can exploit AI’s confident hallucinations; defenders must be trained skeptics. - Education: Curricula should pivot toward philosophy, logic, and process exposition. - Funding models: Support should prioritize programs that deepen human expertise, not just AI deployment. --- ### 🧠 Learning Styles Assessment - Reading level estimate: Approximately upper high-school to early college (Flesch–Kincaid ~11–13), given philosophical references and abstract reasoning. #### 👀 Visual learners - Present: Conceptual metaphors (pilot/autopilot, junior ladder) that can be visualized. - Improve: Add diagrams of “junior to senior” pipelines, decision trees for AI verification, and before/after examples of AI output vs refined human output. #### 🔊 Aural learners - Present: None explicitly (text article). - Improve: Provide a short audio narration or podcast-style discussion of the essay. #### 🗣️ Verbal learners - Present: Rich analogies, philosophical quotes, precise distinctions (thinking vs knowing). - Improve: Glossary of key terms and a short story-based case study illustrating AI failure. #### 🧩 Kinesthetic learners - Present: Implicit encouragement to “debug,” but no concrete tasks. - Improve: Add hands-on exercises where readers must break down and test AI-generated outputs. #### 🧠 Logical learners - Present: Clear argument structure, causal chains, and conceptual models (Descartes reflex, junior paradox). - Improve: Simple tables comparing “average vs excellent worker” outcomes and explicit step-by-step reasoning checklists. #### 🧑‍🤝‍🧑 Social learners - Present: Context of school councils, workplace dynamics, and team-based implications. - Improve: Discussion prompts for classrooms and teams, plus guided debate formats. #### 🧘‍♀️ Solitary learners - Present: Reflective tone, questions about personal responsibility. - Improve: Journaling prompts about one’s own AI use and comfort with uncertainty. #### 🌿 Naturalistic learners - Present: Minimal; mostly abstract and workplace-focused. - Improve: Use ecosystem analogies (e.g., removing junior “seedlings” from a forest of expertise) and discuss sustainability of talent pipelines. --- ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Standard blog structure likely includes headings and paragraphs. - Improve: Ensure proper heading hierarchy, descriptive link text, and language attributes. #### Keyboard navigation - Present: Typical blog layouts are navigable via keyboard. - Improve: Confirm focus indicators, skip links, and logical tab order. #### Contrast and readability - Present: Depends on site theme; WordPress defaults often adequate. - Improve: Verify contrast ratios, font size, and line spacing for long-form reading. #### Text resizing and reflow - Present: Responsive layouts usually support zoom and mobile. - Improve: Check reflow at 200% zoom and support for reduced motion preferences. #### Media and captions - Present: Primarily text; a single external link to Aguilera’s piece. - Improve: If media are added later (images, embeds), provide captions and alt text. #### Forms and interactions - Present: Likely minimal (comments, subscribe). - Improve: Ensure labeled fields, clear errors, and large hit areas. #### Standards reference - Map any gaps to WCAG 2.2 AA, especially 1.4.x (contrast), 2.1.x (keyboard), 1.3.1 (info and relationships). --- ### 🌍 DEIB Assessment #### Cultural breadth - Present: References European philosophers and a Dutch educational context. - Improve: Add examples from non-Western philosophies and different labor markets. #### Inclusive language - Present: Generally neutral, professional language. - Improve: Clarify jargon (e.g., “Gymnasium,” “AI council”) and maintain plain-language summaries. #### Neurodiversity support - Present: Clear sections and headings. - Improve: Further chunk content, add summaries per section, and avoid overly long paragraphs. #### Representation - Present: Focuses on roles (junior/senior, educator/manager) rather than identities. - Improve: Include scenarios showing diverse backgrounds and pathways into tech. #### Psychological safety - Present: Some “uncomfortable truths,” but framed constructively. - Improve: Offer explicit reassurance that skills can be learned and suggest supportive learning environments. --- ### ✅ Logical Validity #### Strong reasoning - The argument that AI commoditizes average output is inductive, supported by examples (essays, code, plans) and consistent with observed tool capabilities. - The “junior paradox” offers a causal chain: removing junior work → fewer opportunities to build deep expertise; this aligns with apprenticeship models across fields. - The Descartes reflex as “intellectual self-defense” is an analogical argument connecting debugging, language parsing, and AI verification; the analogy is tight and intuitive. #### Fallacies - Potential overgeneralization: “average output has no market value anymore” is rhetorically strong but could be softened (“rapidly declining market value”) to acknowledge domains where human average still matters. - Implicit slippery slope: strong AI reliance “removes the very path that made senior developers possible”; could be clarified as a risk contingent on not redesigning training pathways. #### Tone of voice - Tone is confident but reflective, mixing philosophical seriousness with pragmatic workplace concern. - Uses vivid metaphors (“passenger who doesn’t know how to land the plane,” “AI will confidently sign off on anything”) to make abstract points concrete without becoming flippant. --- ### 🛡️ DARVO Analysis - Deny: The piece does not deny AI risks; it foregrounds them. Nor does it deny human responsibility—quite the opposite. - Attack: There is no attack on individuals; criticism targets behaviors (over-trusting AI, skipping the hard work) and market dynamics. - Reverse victim/offender: The author does not cast AI or elites as victims; instead, he emphasizes human agency and responsibility. - Overall: The rhetoric encourages ownership and critical thinking, not defensiveness or blame-shifting. --- ## 🎯 Conclusion AI’s polished confidence can seduce us into outsourcing thinking, but the article argues that genuine value now lies in judgment, decomposition, and the courage to say “I don’t know yet.” As AI raises the floor of average work, individuals, educators, and organizations must redesign learning and hiring around these deeper skills—becoming pilots who supervise powerful tools, not passengers carried along by convincing but unexamined answers. --- ## 🧱 Refactor Appendix (append after 🎯 Conclusion) ### 📚 Literature Split Pad - Purpose: Stage Konik CEE units that become per-source Literature notes. - Split level: **H3**. - Include verbatim quotes under 25 words and a [view highlight] link. ### [H1] 5–10-word present-tense claim > [!quote] [Why healthy doubt beats AI confidence theater](https://joost.blog/healthy-doubt/) by [[Joost de Valk]] via [[joost.blog]]. Published 2026-02-21. > “AI will confidently sign off on anything.” Two lines: context and use. Cite data if relevant. [view highlight](https://joost.blog/healthy-doubt/#H1) #xref #articleseed ID:rw_[MISSING] ### [H2] 5–10-word present-tense claim > [!quote] [Why healthy doubt beats AI confidence theater](https://joost.blog/healthy-doubt/) > “AI makes the average worker obsolete, but the excellent worker irreplaceable.” Two lines: context and use. [view highlight](https://joost.blog/healthy-doubt/#H2) #xref ID:rw_[MISSING] --- ### 🔖 Atomic Split Pad (Zettel Candidates) - Purpose: Promote reusable claims into Atomic notes. - Split level: **H4**. - Each item: one idea, present tense, specific, testable. #### [Z1] AI confidence requires human skepticism to be safe - Scope/condition in one line. - Pointer to page/time or dataset. - Links: [[AI-Judgment]] [[Intellectual-Self-Defense]] #### [Z2] Junior work remains necessary in an AI era - Scope/condition in one line. - Pointer to page/time or dataset. - Links: [[Apprenticeship-Models]] [[AI-Workforce]] --- ### ✏️ Fleeting Queue (Extract by Selection) - Purpose: Capture quick actions or sparks for Fleeting notes. - Extraction: **Select one line** → “Refactor: Extract selection to new note (content only).” - Keep each line short and self-contained. - [F1] Design a workshop on the “Descartes reflex” using AI-generated bugs — context/https://joost.blog/healthy-doubt/ #fleeting #inbox - [F2] Draft hiring rubric that scores AI judgment over output volume — context/https://joost.blog/healthy-doubt/ #fleeting #inbox - [F3] Create classroom prompts that force students to explain AI-assisted work — context/https://joost.blog/healthy-doubt/ #fleeting #inbox --- # 🏷️ Title: Why healthy doubt beats AI confidence theater ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Include links cited in or relevant to the article. --- ## 📝 Summary - What happened. - The author argues that modern AI systems project extreme confidence while lacking self-awareness of ignorance, and that this reshapes work, education, and what “junior” means. - Why it matters. - “Average” output is now cheap and abundant; the scarce skill is human judgment, decomposition, and the courage to doubt AI’s polished answers. - Key stakeholders. - Students, educators, junior and senior workers (especially developers), managers/founders, and institutions designing curricula or hiring practices. - Risks and tradeoffs. - Over-reliance on AI erodes pathways to senior expertise, rewards shallow “prompting” over deep thinking, and encourages outsourcing responsibility to machines. - Bottom line. - AI makes average workers obsolete and excellent workers irreplaceable; cultivating skepticism, philosophical thinking, and “pilot not passenger” skills is now essential. --- ## 📖 Main Content - The author contrasts AI’s “confidence theater” with philosophical traditions of doubt: Socrates’ “I know that I know nothing,” Arendt’s distinction between thinking and knowing, and Kant’s “Sapere aude.” - AI is framed as a “knowing machine” that can generate polished answers but cannot tell when it’s wrong; hallucinations and truths are delivered with the same poise. - Because AI can now produce passable essays, code, and plans instantly, “average output has no market value anymore.” Human value shifts from production to evaluation and judgment. - The “junior paradox”: historically, people became senior by doing large amounts of junior work. If AI does that work, future seniors have no training ground, yet entry-level expectations are rising (“you can do what AI does plus catch where it fails”). - The “pilot vs passenger” model: effective professionals use AI like autopilot—understanding and supervising it—rather than riding as passengers who are helpless when it fails. - The “Descartes reflex” is introduced as the habit of decomposing and verifying complex outputs, akin to debugging or parsing Latin. Classical education is reframed as training for AI-era cognitive skills. - Kant’s “Sapere aude” is updated: instead of church and king, we must stop outsourcing thinking to “algorithm and model.” You may outsource work to AI, but not responsibility. - Concrete recommendations follow for educators (grade process, not product), managers (hire for judgment), entrants to the workforce (embrace “I don’t know yet”), and everyone (treat philosophy as cognitive self-defense). - The essay closes with the “uncomfortable truth”: AI commoditizes average workers but amplifies the value of those with deep, hard-won understanding. --- ### 🗒️ Notes - “AI will confidently sign off on anything. The question is whether you will too.” — sets up the central tension between AI’s confidence and human responsibility. - The author’s context is advising a Dutch gymnasium’s AI council, tying AI’s impact on the workforce back to classical education in Latin, Greek, and philosophy. - “Ask Claude, Gemini, or ChatGPT anything and you’ll get a polished, articulate, supremely confident answer. Even when it’s wrong.” — highlights the confidence problem. - Arendt’s thinking vs knowing: AI excels at producing answers but cannot “examine those answers, sit with uncertainty, [or] refuse to stop at the surface.” - “We’ve built machines that perform certainty, and we’re raising a generation that’s learning to mistake that performance for truth.” — core societal risk. - “Average output has no market value anymore.” Passable essays, working code, basic analysis: all now cheap commodities generated by AI. - Human edge: “It can’t feel that something is off. It can’t exercise judgment born from years of getting things wrong and learning why.” - Junior paradox: seniority historically requires years of “junior work” (bad code, many parsed sentences). AI now erodes that ladder by doing junior work itself. - “Entry-level now means you can do what AI does plus catch where it fails.” — a new, higher bar for juniors. - Pilot vs passenger: real pilots understand and oversee autopilot; good workers must supervise AI like a powerful but fallible intern. - “The market is firing passengers and hiring pilots… eliminating executors and desperately seeking directors.” — preference for architects over pure executors. - Descartes reflex: break problems into small pieces, verify each, then rebuild — the debugging mindset AI threatens by making outputs “look” right. - Classical education (e.g., parsing Latin) is reframed as practice in decomposition and reconstruction—exactly what’s needed to interrogate AI output. - Kant’s Sapere aude: stop outsourcing thinking to authority; today’s authority is “algorithm and model,” not church and king. - “You can use AI to outsource the work. You can’t outsource the responsibility.” — ethical boundary of AI use. - Educator advice: stop grading only the final product; interrogate the process with questions like “Why this word and not that one?” and “How do you know this is right?” - Manager advice: “Hire for judgment, not output volume… The person who can tell you why the AI’s answer is wrong is worth ten people who can generate AI answers quickly.” - Worker advice: your value lies “in the gap between ‘this looks right’ and ‘this is right’” and the willingness to say “I don’t know yet.” - Philosophy becomes “cognitive self-defense”: deconstructing arguments, questioning assumptions, and judging under uncertainty. - Final claim: “AI makes the average worker obsolete, but the excellent worker irreplaceable… The wind is in the sails. Humans still set the course.” --- ## 📚 References - Christina Aguilera, “AI Confidence Theater” (LinkedIn leadership piece) - Immanuel Kant, “Answering the Question: What Is Enlightenment?” (1784) - Hannah Arendt, works distinguishing thinking vs knowing (esp. “The Life of the Mind”) - Aristotle, “Nicomachean Ethics” (practice, virtue, habituation) - René Descartes, “Discourse on Method” (decomposition of problems) - Joost de Valk, “The death of Code Copyright and the rise of the Architect” (referenced prior post) - Original article: “Why healthy doubt beats AI confidence theater” — https://joost.blog/healthy-doubt/ --- ## 🗣️ Paraphrase - Paragraph A paraphrase. - Modern AI systems answer any question with polished confidence, regardless of correctness, and society risks mistaking that performance for genuine understanding. - Paragraph B paraphrase. - As AI automates mediocre work, human value shifts toward critical thinking, skepticism, and the ability to dissect and correct AI output—skills long cultivated by philosophy and classical education. --- ## 📰 Newsletter AI tools can now produce “good enough” content on command—but Joost de Valk argues that’s exactly why healthy doubt matters more than ever. In his latest essay, he connects classical philosophy to today’s AI-powered workplace, showing why WordPress professionals must be pilots, not passengers, when using tools like GitHub Copilot or content generators. Worth a read if you’re rethinking workflows, training juniors, or designing AI-assisted editorial processes in WordPress. --- ## ✅ Action Items #### DevRel - Prioritize talks and demos that show debugging and verifying AI output, not just generating it. - Create workshops teaching “Descartes reflex” skills: code review, reasoning checks, and prompt-plus-judgment patterns. - Produce example repos and quickstarts that pair AI-assisted code with human review checklists. - Run campaigns contrasting “average AI output” vs curated, explained solutions to highlight human value. - Facilitate community office hours focused on reading and critiquing AI-generated patches. #### OSPO - Draft guidance clarifying that AI-generated contributions still require human accountability and review. - Update contribution policies to document provenance and review of AI-assisted code. - Integrate security scanning and license checks that account for AI-suggested dependencies. - Maintain inventories (SBOMs) that note where AI-assisted changes enter critical components. - Educate contributors and maintainers on risks of blindly accepting AI patches and the need for verification. #### Community - Offer “architecture clinics” where maintainers walk through how they evaluate AI-generated proposals. - Design contributor onboarding that teaches how to interrogate AI output instead of pasting it verbatim. - Encourage sponsors to fund mentorship for juniors learning real debugging and design, not just prompting. - Coordinate cross-project incident drills involving AI-induced bugs and their remediation. #### WordPress - Core: Emphasize code review standards for AI-assisted patches and require clear rationale in tickets. - Core Performance: Benchmark AI-generated optimizations and validate with profiling before merging. - Core AI: Establish responsible-use guidelines, evaluation sets, and failure-mode exemplars. - Core Program Team: Add regular reviews of how AI affects contributor pathways (especially juniors). - Mobile: Document where AI helps with boilerplate while ensuring humans own UX and edge cases. - Playground: Provide sandboxes for experimenting with AI-generated code in a safe environment. - Community: Include “AI skepticism” modules in meetup kits and WordCamp talks. - Training: Build lesson plans on reading, testing, and refactoring AI-generated code/content. - Marketing: Position WordPress AI features as copilots that still require human editors. - Support: Capture common AI-related support issues and encode them into docs and training. - Design: Audit AI-generated UI ideas with human-led usability testing. - Documentation: Add examples explaining how to verify AI-written snippets and queries. - Photos: Clarify policies around AI-generated media and attribution. - TV: Encourage talks on AI literacy, doubt, and responsible usage. - Polyglots: Evaluate AI-assisted translations with human reviewers, documenting known failure modes. - Accessibility: Ensure AI-generated interfaces and content are checked against WCAG. - Meta: Provide tooling to flag suspiciously generic (likely AI) contributions for extra review. - CLI: Offer safe scaffolds but require explicit human review steps for generated commands. - Hosting: Share best practices for AI-assisted configuration, logging, and debugging. - Tide: Consider how AI-generated plugins/themes affect scoring and verification. - Openverse: Clarify inclusion rules and labeling for AI-generated media. - Themes: Mentor authors on using AI for drafts while maintaining human-led design systems. - Plugins: Strengthen security reviews where AI may have introduced subtle flaws. #### The WPOCC - Programs: Fund fellowships teaching critical AI use and architectural thinking to maintainers. - Projects: Run documentation sprints on AI best practices and governance. - Funding experiments: Pilot grants for tools that make AI output more inspectable and auditable. ### FAIR - Encourage local evaluation and metadata to track AI involvement in code and content. - Publish governance artifacts outlining acceptable AI use and transparent funding of AI tooling. --- ## 🧪 Hypotheses - H1: Workers trained in decomposition and philosophical reasoning will detect AI errors more reliably than those trained only in tool usage. - H2: Organizations that hire for judgment over output volume will experience fewer AI-induced production incidents. - H3: Curricula that assess students on process, not just product, will better prepare them for AI-saturated workplaces. - H4: Over-reliance on AI for junior tasks will reduce the future supply of true senior experts. - H5: Explicitly teaching “pilot vs passenger” mental models will improve responsible AI usage. --- ### 📊 Methodology - Run controlled experiments comparing groups with and without explicit training in the “Descartes reflex” on tasks involving flawed AI outputs. - Conduct longitudinal studies across companies adopting AI, tracking incident rates vs hiring practices (judgment-focus vs volume-focus). - Compare outcomes for students in programs with process-based grading vs traditional essay grading in AI-rich environments. - Analyze career trajectories in teams heavily automating junior tasks vs those preserving manual practice opportunities. - Use surveys and behavioral assessments to measure “pilot vs passenger” attitudes and correlate them with real-world error detection. --- ### ❓ Questions - How can we design scalable training pathways that preserve junior learning while leveraging AI? - What signals reliably indicate human judgment in hiring beyond portfolio volume? - How should educational accreditation adapt in a world where AI can write any assignment? - Which parts of software and content pipelines should remain strictly human-signed? - How can open-source communities transparently track AI involvement without deterring contributors? --- ### 🔮 Future Research - Develop metrics for “judgment density” in roles and organizations. - Explore tools that make AI reasoning chains visible and critique-able. - Investigate hybrid human–AI apprenticeship models that maintain deep practice. - Study long-term cognitive impacts of frequent AI reliance on problem-solving. - Examine cross-cultural differences in trust of AI authority and how they affect adoption. --- ### 💡 Implications - Practice: Professionals must cultivate debugging, decomposition, and argument analysis as core skills alongside tool fluency. - Policy: Institutions may need guidelines assigning final responsibility for AI-assisted decisions. - Security: Attackers can exploit AI’s confident hallucinations; defenders must be trained skeptics. - Education: Curricula should pivot toward philosophy, logic, and process exposition. - Funding models: Support should prioritize programs that deepen human expertise, not just AI deployment. --- ### 🧠 Learning Styles Assessment - Reading level estimate: Approximately upper high-school to early college (Flesch–Kincaid ~11–13), given philosophical references and abstract reasoning. #### 👀 Visual learners - Present: Conceptual metaphors (pilot/autopilot, junior ladder) that can be visualized. - Improve: Add diagrams of “junior to senior” pipelines, decision trees for AI verification, and before/after examples of AI output vs refined human output. #### 🔊 Aural learners - Present: None explicitly (text article). - Improve: Provide a short audio narration or podcast-style discussion of the essay. #### 🗣️ Verbal learners - Present: Rich analogies, philosophical quotes, precise distinctions (thinking vs knowing). - Improve: Glossary of key terms and a short story-based case study illustrating AI failure. #### 🧩 Kinesthetic learners - Present: Implicit encouragement to “debug,” but no concrete tasks. - Improve: Add hands-on exercises where readers must break down and test AI-generated outputs. #### 🧠 Logical learners - Present: Clear argument structure, causal chains, and conceptual models (Descartes reflex, junior paradox). - Improve: Simple tables comparing “average vs excellent worker” outcomes and explicit step-by-step reasoning checklists. #### 🧑‍🤝‍🧑 Social learners - Present: Context of school councils, workplace dynamics, and team-based implications. - Improve: Discussion prompts for classrooms and teams, plus guided debate formats. #### 🧘‍♀️ Solitary learners - Present: Reflective tone, questions about personal responsibility. - Improve: Journaling prompts about one’s own AI use and comfort with uncertainty. #### 🌿 Naturalistic learners - Present: Minimal; mostly abstract and workplace-focused. - Improve: Use ecosystem analogies (e.g., removing junior “seedlings” from a forest of expertise) and discuss sustainability of talent pipelines. --- ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Standard blog structure likely includes headings and paragraphs. - Improve: Ensure proper heading hierarchy, descriptive link text, and language attributes. #### Keyboard navigation - Present: Typical blog layouts are navigable via keyboard. - Improve: Confirm focus indicators, skip links, and logical tab order. #### Contrast and readability - Present: Depends on site theme; WordPress defaults often adequate. - Improve: Verify contrast ratios, font size, and line spacing for long-form reading. #### Text resizing and reflow - Present: Responsive layouts usually support zoom and mobile. - Improve: Check reflow at 200% zoom and support for reduced motion preferences. #### Media and captions - Present: Primarily text; a single external link to Aguilera’s piece. - Improve: If media are added later (images, embeds), provide captions and alt text. #### Forms and interactions - Present: Likely minimal (comments, subscribe). - Improve: Ensure labeled fields, clear errors, and large hit areas. #### Standards reference - Map any gaps to WCAG 2.2 AA, especially 1.4.x (contrast), 2.1.x (keyboard), 1.3.1 (info and relationships). --- ### 🌍 DEIB Assessment #### Cultural breadth - Present: References European philosophers and a Dutch educational context. - Improve: Add examples from non-Western philosophies and different labor markets. #### Inclusive language - Present: Generally neutral, professional language. - Improve: Clarify jargon (e.g., “Gymnasium,” “AI council”) and maintain plain-language summaries. #### Neurodiversity support - Present: Clear sections and headings. - Improve: Further chunk content, add summaries per section, and avoid overly long paragraphs. #### Representation - Present: Focuses on roles (junior/senior, educator/manager) rather than identities. - Improve: Include scenarios showing diverse backgrounds and pathways into tech. #### Psychological safety - Present: Some “uncomfortable truths,” but framed constructively. - Improve: Offer explicit reassurance that skills can be learned and suggest supportive learning environments. --- ### ✅ Logical Validity #### Strong reasoning - The argument that AI commoditizes average output is inductive, supported by examples (essays, code, plans) and consistent with observed tool capabilities. - The “junior paradox” offers a causal chain: removing junior work → fewer opportunities to build deep expertise; this aligns with apprenticeship models across fields. - The Descartes reflex as “intellectual self-defense” is an analogical argument connecting debugging, language parsing, and AI verification; the analogy is tight and intuitive. #### Fallacies - Potential overgeneralization: “average output has no market value anymore” is rhetorically strong but could be softened (“rapidly declining market value”) to acknowledge domains where human average still matters. - Implicit slippery slope: strong AI reliance “removes the very path that made senior developers possible”; could be clarified as a risk contingent on not redesigning training pathways. #### Tone of voice - Tone is confident but reflective, mixing philosophical seriousness with pragmatic workplace concern. - Uses vivid metaphors (“passenger who doesn’t know how to land the plane,” “AI will confidently sign off on anything”) to make abstract points concrete without becoming flippant. --- ### 🛡️ DARVO Analysis - Deny: The piece does not deny AI risks; it foregrounds them. Nor does it deny human responsibility—quite the opposite. - Attack: There is no attack on individuals; criticism targets behaviors (over-trusting AI, skipping the hard work) and market dynamics. - Reverse victim/offender: The author does not cast AI or elites as victims; instead, he emphasizes human agency and responsibility. - Overall: The rhetoric encourages ownership and critical thinking, not defensiveness or blame-shifting. --- ## 🎯 Conclusion AI’s polished confidence can seduce us into outsourcing thinking, but the article argues that genuine value now lies in judgment, decomposition, and the courage to say “I don’t know yet.” As AI raises the floor of average work, individuals, educators, and organizations must redesign learning and hiring around these deeper skills—becoming pilots who supervise powerful tools, not passengers carried along by convincing but unexamined answers. --- ## 🧱 Refactor Appendix (append after 🎯 Conclusion) ### 📚 Literature Split Pad - Purpose: Stage Konik CEE units that become per-source Literature notes. - Split level: **H3**. - Include verbatim quotes under 25 words and a [view highlight] link. ### [H1] 5–10-word present-tense claim > [!quote] [Why healthy doubt beats AI confidence theater](https://joost.blog/healthy-doubt/) by [[Joost de Valk]] via [[joost.blog]]. Published 2026-02-21. > “AI will confidently sign off on anything.” Two lines: context and use. Cite data if relevant. [view highlight](https://joost.blog/healthy-doubt/#H1) #xref #articleseed ID:rw_[MISSING] ### [H2] 5–10-word present-tense claim > [!quote] [Why healthy doubt beats AI confidence theater](https://joost.blog/healthy-doubt/) > “AI makes the average worker obsolete, but the excellent worker irreplaceable.” Two lines: context and use. [view highlight](https://joost.blog/healthy-doubt/#H2) #xref ID:rw_[MISSING] --- ### 🔖 Atomic Split Pad (Zettel Candidates) - Purpose: Promote reusable claims into Atomic notes. - Split level: **H4**. - Each item: one idea, present tense, specific, testable. #### [Z1] AI confidence requires human skepticism to be safe - Scope/condition in one line. - Pointer to page/time or dataset. - Links: [[AI-Judgment]] [[Intellectual-Self-Defense]] #### [Z2] Junior work remains necessary in an AI era - Scope/condition in one line. - Pointer to page/time or dataset. - Links: [[Apprenticeship-Models]] [[AI-Workforce]] --- ### ✏️ Fleeting Queue (Extract by Selection) - Purpose: Capture quick actions or sparks for Fleeting notes. - Extraction: **Select one line** → “Refactor: Extract selection to new note (content only).” - Keep each line short and self-contained. - [F1] Design a workshop on the “Descartes reflex” using AI-generated bugs — context/https://joost.blog/healthy-doubt/ #fleeting #inbox - [F2] Draft hiring rubric that scores AI judgment over output volume — context/https://joost.blog/healthy-doubt/ #fleeting #inbox - [F3] Create classroom prompts that force students to explain AI-assisted work — context/https://joost.blog/healthy-doubt/ #fleeting #inbox --- <!-- deep-read:sections end -->