Almost Timely News: 🗞️ the Biggest Problem With AI Today  🔗 **Source:** [Almost Timely News: 🗞️ the Biggest Problem With AI Today](https://www.christopherspenn.com/2026/07/almost-timely-news-%f0%9f%97%9e%ef%b8%8f-the-biggest-problem-with-ai-today-2026-07-05/) > [!note]+ Document Notes > tags: [executive-function, metacognition, critical-thinking, agentic-ai, deskilling] > author: Christopher S. Penn > source: Christopher S. Penn / Almost Timely News > reference: https://www.christopherspenn.com/2026/07/almost-timely-news-%f0%9f%97%9e%ef%b8%8f-the-biggest-problem-with-ai-today-2026-07-05/ > date created: 2026-07-05 13:10 > > # 🏷️ Title: Almost Timely News: 🗞️ The Biggest Problem with AI Today (2026-07-05) > <!-- 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 - [The Biggest Problem with AI Today — Christopher S. Penn](https://www.christopherspenn.com/2026/07/almost-timely-news-%f0%9f%97%9e%ef%b8%8f-the-biggest-problem-with-ai-today-2026-07-05/) - [The Problem with Streaming — and the Case for Physical Media (Tom Rizzuto, TED)](https://www.youtube.com/watch?v=yUzvXA5HmtE) - Trust Insights CASINO research framework (referenced in-text; no direct link supplied) - NotebookLM + NotebookLM CLI tool (source-analysis toolchain referenced in-text) ## 📝 Summary - **What happened.** Penn argues the biggest problem in AI today is not cost, sustainability, or ethics — those are symptoms. The root problem is that we've abdicated our *executive function* to AI: "nobody's making decisions." He built the case by loading 180 days of 40+ marketing/business/AI subreddits (~800,000 words) into NotebookLM, wiring it to Claude Code, and comparing Reddit's top AI pain points against his own 2026 newsletters. Claude surfaced 10 pain-point topics; Penn reframes all 10 as one failure: we've stopped thinking about thinking. - **Why it matters.** Executive function — which he buckets as PODS (Plan, Organize, Decide, Solve) — is the seat of accountability, skill, and professional value. Offloading it produces job loss, deskilling, and dissatisfaction, and once a skill goes rusty it becomes cheaper to replace the person. - **Key stakeholders.** Marketers, analysts, agency owners, knowledge workers, and the vendors selling AI tooling and "AI visibility" measurement; secondarily the platforms whose default-to-expensive-model design shapes habits. - **Risks and tradeoffs.** Approving machine output without leading it removes the human decision-maker; the counter-move (lead first, force the machine to educate you) costs time and effort up front. - **Bottom line.** Two-question gate for every task that matters: does AI make the output better, and does AI make *me* better? Must be yes to both. ## 📖 Main Content - **The thesis inversion.** Penn opens by listing the usual "biggest problem" candidates (token cost, electricity/water sustainability, ethics of AI-cramming) and dismisses them as downstream symptoms. The upstream cause is abdicated executive function — a governance and cognition failure, not a resource or ethics failure. - **How the topic was generated.** He describes his method transparently: a NotebookLM notebook seeded with 180 days of 40+ subreddits, connected to Claude Code via the NotebookLM CLI tool (framed as the most token-efficient bridge), with his year-to-date newsletters in an input folder. He asked Claude to contrast his coverage against Reddit's hardest AI problems. Claude returned 10 candidate topics from ~800,000 words. - **The self-aware override.** Claude pushed hard toward a "measurement is broken in marketing and AI" framing. Penn deliberately overrode that nudge, noting measurement is a *trailing* indicator, not a leading one — and that broken measurement is itself a symptom of abdicated executive function. This override is itself an act of exercising executive function, which is the article's point performed rather than merely stated. - **Executive function defined (PODS).** Plan (chart a path from here to a future goal), Organize (make sense of what you have), Decide (make choices about it), Solve (fix the problems you have). He equates this with critical thinking, which he says is badly named: it's really *metacognition* — thinking about thinking, asking "is this the best way?" and "how could I do this better?" - **The sentience framing.** He claims executive function defines sentience across creatures (crows bending wire tools, cats optically measuring a jump). He then concedes that well-prompted AI, given frameworks, harnesses, and data, can plan/organize/decide/solve better than humans at most *language-based* tasks — "and therein lies the actual problem." The capability is real; the danger is ceding the function wholesale. - **The tale of the tape (10 threads → 1 root).** Each Reddit pain point is re-read as a metacognition failure: AI visibility metrics are invented because stakeholders prefer a confident wrong number to honest uncertainty; agentic oversight degrades so people just click OK (Claude's "dangerously skip permissions," Qwen's "YOLO mode"); deployment is declared without measuring impact (29% of companies see significant ROI while individuals claim 5x productivity — "the math doesn't math"); 40–60% of budget wasted because everyone accepts the default (often most-expensive) model, plausibly by habit-forming design; "AI is a rental" — you own little in closed-weights SaaS, echoing Anthropic's Fable 5 export-control wobble; sycophancy turns synthetic focus groups into confirmation-bias machines; AI detectors falsely flag human work (1 in 7 in Penn's own test); companies hollow out task-by-task with nobody deciding; tokenmaxxing (Meta measuring productivity by token spend); marketers as unpaid trainers for platforms competing for their jobs. - **The antidote.** For every task that matters, lead first and force the machine to educate you. His own practice: dictate his own analysis of client data aloud, transcribe it, *then* have AI review it, quiz him, and pull more out of him; ask AI to find learning resources within his expertise (he had it teach him one of three analytical techniques he didn't know); outside his expertise, use agentic deep research (Claude Code, OpenCode, CASINO framework), then build a quality checklist and learn the domain himself via infographics/podcast summaries. The value he sells is the lens and expertise, not being "a meat-based interface to an LLM." - **Contested / uncertain claims to flag.** The 29%-vs-5x juxtaposition compares different units (org-level ROI vs self-reported individual productivity) and is rhetorical, not arithmetic. The generalization from Reddit verbatims to "most people" is inductive and unquantified. The "executive function defines our sentience" line leans on an appeal-to-nature framing. The ~90% purchase-intent-emulation figure is asserted without citation. ### 🗒️ Notes - Frame: the "biggest problem" question is a bait-and-switch — cost/sustainability/ethics are named only to be demoted to symptoms of a cognition failure. - Core equation: critical thinking = metacognition = thinking about thinking; PODS (Plan, Organize, Decide, Solve) is the operational handle. - Method matters as much as message: NotebookLM (180 days × 40+ subreddits, ~800k words) → Claude Code via NotebookLM CLI → contrast with his own newsletters → 10 topics. - The 10 topics: (1) AI visibility measurement invented; (2) agentic oversight degrading (skip-permissions / YOLO mode); (3) deployment declared without impact (29% ROI vs 5x claims); (4) 40–60% budget wasted on default expensive models (habit-formation-by-design); (5) AI is a rental (own nothing in closed-weights SaaS; Fable 5 export-control scare); (6) sycophancy corrupts synthetic focus groups (confirmation bias); (7) AI detectors don't work (1-in-7 false flag on human text); (8) companies hollowing out task-by-task ("nobody decided this"); (9) tokenmaxxing (Meta measuring by token spend); (10) marketers as unpaid training labor. - Recurring diagnostic sentence across all 10: "they are not thinking about their thinking." - Accountability argument: when the machine decides, there is no one accountable; approval ≠ decision-making. - Pride/effort analogies: child's bad-but-treasured art (value = effort, not quality); store-bought cake and the allergen question (you can't answer what you didn't make). - Ownership analogy (cross-link anchor): SaaS ownership is like Spotify music or Netflix video — you don't own it, but you think you do. Directly parallels Tom Rizzuto's streaming-vs-physical-media argument. - Antidote mechanics: lead → dictate own read of the data → transcribe → AI critiques/quizzes you → AI sources learning material → you level up domain knowledge. Outside expertise: deep research → quality checklist → self-teach. - Value thesis: pay is for the lens and expertise, not for being a copy-paste drone; AI that improves your expertise justifies your continued employment. - Meta-irony worth preserving: Penn overrode Claude's "measurement is broken" nudge — the article's method demonstrates the very executive function it prescribes. - Two-question gate (the takeaway to quote): (1) does AI make the output better? (2) does AI make me better? Yes to both, or you're using it wrong. - Closing warning: offloaded skills go rusty; rusty enough → cheaper to replace. ## 📚 References - [The Biggest Problem with AI Today — Christopher S. Penn](https://www.christopherspenn.com/2026/07/almost-timely-news-%f0%9f%97%9e%ef%b8%8f-the-biggest-problem-with-ai-today-2026-07-05/) - [The Problem with Streaming — and the Case for Physical Media (Tom Rizzuto, TED)](https://www.youtube.com/watch?v=yUzvXA5HmtE) - Trust Insights CASINO research framework — referenced in-text as the agentic-research scaffold (no primary link in source). - NotebookLM (Google) and the NotebookLM command-line tool — the analysis toolchain described in Part 1. ## 🗣️ Paraphrase - The problems people obsess over in AI — what it costs, what it burns, how ethically it's deployed — are downstream. The real trouble is that we've stopped making decisions and handed the deciding to the machine. - Penn didn't just assert this; he mined 180 days of 40-plus subreddits through NotebookLM and Claude Code, got ten "hardest AI problems," and noticed the AI itself wanted to steer him toward "measurement is broken." He said no — measurement is a lagging symptom. - Executive function, his PODS (plan, organize, decide, solve), is the machinery of thinking about your own thinking. Every one of the ten Reddit complaints is really that machinery switched off. - Modern AI can do those four things better than us on language tasks, which is exactly why it's tempting to surrender them — and surrendering them is the trap. - The fix isn't abstinence. For anything that matters, do the thinking first, then make the tool teach you and improve you. If the AI both improves the output and improves you, keep using it. If not, you're deskilling yourself toward replaceability. ## 📰 Newsletter - Christopher Penn's latest *Almost Timely* is a sharp gut-check for anyone shipping AI-assisted work: the biggest problem isn't the token bill, it's that "nobody's making decisions." He calls it abdicating your executive function — PODS: Plan, Organize, Decide, Solve — and reframes ten of Reddit's loudest AI gripes as one quiet failure to think about thinking. - The WordPress and DevRel angle writes itself: the same "just click OK" reflex that degrades agentic oversight is the reflex that lets a generated tutorial, changelog, or docs page ship without a human who actually understands it. Penn's antidote is a keeper — lead the task first, then make the tool teach you — and his two-question test is a clean editorial gate: does AI make the output better, and does it make *me* better? - Pairs beautifully with Tom Rizzuto's TED talk on streaming vs. physical media. Rizzuto says don't outsource ownership of your media; Penn says don't outsource ownership of your thinking. Same spine, different shelf. Worth 15 minutes and a team discussion. ## ✅ Action Items #### DevRel - Developer advocacy: build a talk/demo on "lead-first AI" — show the workflow where you dictate your own analysis, then have the model critique and quiz you, with a measurable before/after skill outcome. - Education: run a workshop on not outsourcing judgment — teach the two-question gate (better output *and* better practitioner) as a checkpoint in any AI-assisted lesson. - Developer experience: ship a "force the machine to educate you" quickstart/template — a prompt-and-checklist scaffold that turns tool output into a learning loop, not a copy-paste drone. - Developer marketing: content series contrasting "meat-based interface to an LLM" vs. expert-in-the-loop; attribute engagement to the metacognition framing. - Community management: seed a discussion prompt asking members where they've "hit OK" without deciding; moderate toward reflective sharing. - Internal/external/cross-project: adopt the PODS check as a review-gate norm across advocacy work; feed findings back to Core AI. #### OSPO - Policy & Governance: add a "human-decision-of-record" clause to contribution policy for AI-assisted PRs — name who decided, not just who approved. - License Compliance: Not applicable — source raises no license-scanning concern. - Security Management: flag "dangerously skip permissions" / YOLO-mode agent patterns in tooling guidance as an oversight-degradation risk. - Software Inventory: Not applicable — no SBOM angle in this source. - Developer Support: offer consultation hours on building domain quality checklists before adopting agentic research output. - Community Engagement: Not applicable — no upstream-funding angle here. - Strategic Alignment: report "AI deployed" vs. "AI delivering results" as distinct KPIs, per Penn's using-vs-getting-results distinction. #### Community - Maintainers: add a release-hygiene note — generated release notes need a maintainer who can answer the first unscripted question ("the allergen test"). - Contributors: onboarding path that teaches lead-first AI use so first contributions build skill, not dependence. - Sponsors: Not applicable — no sponsorship dimension in this source. - Ecosystem: Not applicable — no cross-project incident angle here. #### WordPress - Core: Not applicable — no core-triage or release-criteria content in this source. - Core Performance: Not applicable — performance budgets are outside this source's scope. - Core AI: define responsible-AI guardrails around human accountability — require a named decision-maker and an evaluation gate (does the output improve; does the contributor's skill improve) for AI features. - **Core Program Team:** Not applicable — no cross-team coordination ritual arises from this source. - Mobile: Not applicable — no mobile-app concern in this source. - Playground: Not applicable — no Playground/education-module hook here (beyond generic). - Community: Not applicable — covered under the top-level Community subteam above. - Training: build a lesson/learning path on metacognition-first AI use — PODS, the two-question gate, and "have the model teach you the technique you don't know." - Marketing: craft a narrative and editorial calendar around "don't outsource your thinking," and audit AI-assisted marketing output against the two-question gate before publishing. - Support: Not applicable — no support-SLA content in this source. - Design: Not applicable — no UX/design-token angle in this source. - Documentation: adopt an examples-first, human-understood-it standard — no AI-generated doc ships unless an author can answer off-script questions about it (directly counters the "store-bought cake" failure). - Photos: Not applicable — no media-library concern here. - TV: Not applicable — no video-submission concern here. - Polyglots: Not applicable — no localization concern in this source. - Accessibility: Not applicable — source raises no WCAG issue. - Meta: Not applicable — no .org infrastructure concern here. - CLI: Not applicable — no command-coverage concern in this source. - Hosting: Not applicable — no hosting-defaults concern here. - Tide: Not applicable — no plugin-scoring concern in this source. - Openverse: Not applicable — no ingestion/moderation concern here. - Themes: Not applicable — no theme-review concern in this source. - Plugins: Not applicable — no plugin-review concern in this source. #### The WPCC - Programs: Not applicable — no fellowship/microgrant hook in this source. - Projects: Not applicable — no governance-clinic hook here. - Funding experiments: Not applicable — no funding-mechanism content in this source. ### FAIR - Technical independence: use Penn's "AI is a rental" point to argue for local metadata/provenance — you own little in closed-weights SaaS, so keep independence where it counts. - Policies and covenants: Not applicable — no funding-transparency covenant arises directly from this source. - Productized governance artifacts: template a "decision-of-record" scorecard so AI-assisted outputs carry an accountable human owner. ## 🧪 Hypotheses - Teams that require a named human decision-maker on AI-assisted deliverables produce work that survives stakeholder follow-up questions at a higher rate than teams that only require approval. - Practitioners who use the "lead first, then let the model teach you" workflow retain or grow measured domain skill over 6 months, while copy-paste users show measurable skill decay. - Organizations measuring "AI delivering results" separately from "AI deployed" report more accurate ROI than those conflating the two. - Defaulting AI tools to a cheaper model (opt-in to expensive) reduces spend without a proportional drop in task-quality ratings. - The two-question gate (better output AND better practitioner), applied as an editorial checkpoint, lowers the rate of published "workslop." ## 📊 Methodology - Skill-decay study: longitudinal design, two cohorts (lead-first vs. offload-first) on a fixed task type; measure blind-rated task quality and an independent skill assessment at 0/3/6 months. - Accountability study: A/B org policy (named decision-of-record vs. approval-only); outcome = rate of unanswered stakeholder follow-ups and rework hours. - ROI-framing study: survey plus ledger reconciliation across firms; regress reported ROI accuracy on whether "using AI" and "results from AI" are tracked separately. - Model-default experiment: randomize default model (cheap vs. expensive) across matched user groups; measure spend and human quality ratings. - Detector-reliability replication: re-run Penn's false-positive test at larger N across detectors to validate the ~1-in-7 human-flag rate. ## ❓ Questions - What operational signal distinguishes "approved the output" from "actually decided," and can it be audited? - Which tasks genuinely warrant lead-first effort, and which are safe to fully offload without long-run skill cost? - Does forcing the model to teach you scale across a team, or does it depend on an already-expert operator like Penn? - If everyone opts out of default expensive models, how does that reshape vendor subsidization economics? - Where is the line between "AI adds to my insight" and "AI substitutes for it," in practice rather than in principle? ## 🔮 Future Research - Instrumentation for metacognition in workflows: can tools surface *where* a human made a real decision vs. rubber-stamped? - Cross-domain generalization of the ownership analogy: does treating cognition like media ownership (see [[The Problem with Streaming and the Case for Physical Media — Tom Rizzuto TED]]) predict who resists deskilling? - Longitudinal deskilling curves by task type — which skills go rusty fastest under offloading. - Design-ethics study of "habit-formation" defaults (expensive-model-by-default) as a dark pattern. - Evaluation-set design for "does the AI make me better," an under-measured second axis alongside output quality. ### 💡 Implications - **Practice:** Bake a decision-of-record and the two-question gate into review workflows; treat AI as a tutor that quizzes you, not an oracle you approve. - **Policy:** Distinguish "deployed" from "delivering"; require accountable human owners for AI-assisted outputs. - **Security:** Treat "skip permissions"/YOLO agent modes as oversight-degradation risks needing explicit sign-off. - **Education:** Teach metacognition-first AI use (PODS, lead-first, self-teaching loops) as core curriculum, not an add-on. - **Funding models:** "AI is a rental" — budget for the reality that closed-weights SaaS confers little ownership; weigh independence where portability matters. - **Cross-link:** This extends and rhymes with [[The Problem with Streaming and the Case for Physical Media — Tom Rizzuto TED]] on the ownership-and-agency axis: Rizzuto argues against outsourcing ownership of your media; Penn argues against outsourcing ownership of your thinking. Same structural worry (convenient access masking lost control), different object. ### 🧠 Learning Styles Assessment - Reading level estimate: Flesch–Kincaid Grade ~9–10 (conversational long-form; short declaratives interleaved with longer explanatory sentences; low jargon density outside "executive function" and "metacognition"). #### 👀 Visual learners - Present: none — no diagrams, charts, or screenshots in the body. - Improve: a PODS diagram, a "10 threads → 1 root" fishbone, and a before/after of the lead-first workflow. #### 🔊 Aural learners - Present: newsletter is text-only in this body. - Improve: 3–5 minute audio summary or the AI-generated podcast summary Penn himself describes. #### 🗣️ Verbal learners - Present: strong — analogies (child's art, store-bought cake/allergen, Spotify/Netflix ownership), plus defined terms (PODS, metacognition). - Improve: a short glossary and a term-comparison of "using AI" vs. "getting results from AI." #### 🧩 Kinesthetic learners - Present: partial — the dictate-transcribe-critique workflow is a walkthrough. - Improve: a reproducible checklist and a small hands-on exercise applying the two-question gate to one real task. #### 🧠 Logical learners - Present: partial — implicit "symptoms → root cause" structure. - Improve: a decision tree for "lead-first vs. offload," and a table mapping each of the 10 topics to its PODS failure. #### 🧑‍🤝‍🧑 Social learners - Present: light — Reddit verbatims imply community voice. - Improve: a team discussion guide and peer-review rubric built on the two-question gate. #### 🧘‍♀️ Solitary learners - Present: strong — reflective, first-person modeling of self-questioning. - Improve: journaling prompts ("where did I click OK without deciding this week?"). #### 🌿 Naturalistic learners - Present: partial — sustainability named then set aside; sentience/animal examples (crows, cats). - Improve: lifecycle framing of skill (grow/atrophy) as an ecosystem analogy. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: clear heading hierarchy (Parts 1–5), linear reading order. - Improve: ensure any republished version uses real heading elements and descriptive link text (not "here"). #### Keyboard navigation - Present: N/A for prose; depends on the site chrome. - Improve: verify newsletter template has visible focus and logical tab order. #### Contrast and readability - Present: N/A in plain text. - Improve: confirm 4.5:1 body contrast and adequate line-height in the email/web render. #### Text resizing and reflow - Present: prose reflows. - Improve: test at 200% zoom in the email client; avoid fixed-width containers. #### Media and captions - Present: none needed for this text body; the cross-linked TED talk requires captions/transcript. - Improve: provide a transcript for any audio summary; caption the Rizzuto video reference. #### Forms and interactions - Present: N/A — no forms in the body. - Improve: ensure the subscribe/unsubscribe controls have labels and adequate target size. #### Standards reference - Map gaps to WCAG 2.2 AA: 1.4.3 (Contrast), 1.4.10 (Reflow), 1.4.4 (Resize Text), 2.4.4 (Link Purpose), 1.2.x (captions/transcripts for the linked media). ### 🌍 DEIB Assessment #### Cultural breadth - Present: US-centric framing (US holiday weekend, US export controls). - Improve: non-US examples of AI deployment norms and labor context. #### Inclusive language - Present: largely neutral, second-person, jargon defined (PODS, metacognition). - Improve: gloss "tokenmaxxing," "workslop," "GEO," and "synthetic focus groups" for non-specialist readers. #### Neurodiversity support - Present: chunked into Parts, consistent structure, short paragraphs. - Improve: add summary callouts and icon+text signposting; define terms at first use. #### Representation - Present: voices are aggregated Reddit verbatims and the author; limited attributed diversity. - Improve: cite named practitioners across roles/regions rather than anonymous aggregate ranting. #### Psychological safety - Present: respectful, non-blaming tone ("we've," not "you've," mostly). - Improve: frame deskilling risk without fatalism; add resources for workers facing job erosion. ### ✅ Logical Validity #### Strong reasoning - **Inductive synthesis (Part 3, all 10 topics):** aggregating ~800k words of complaints into one root cause is a strong inductive move; the repeated "not thinking about their thinking" refrain gives it a coherent through-line. Strengthen by quantifying how many verbatims actually fit the metacognition frame vs. how many were forced. - **Analogical reasoning (Part 4):** the store-bought-cake/allergen analogy and the child's-art/effort analogy are apt and memorable illustrations of "you can't answer for work you didn't do" and "value tracks effort." Strong because the mapped relation (ownership/effort → accountability) holds; clarify the limits (some outsourced work is fine to not understand deeply). - **Self-referential demonstration (Parts 1 & 3):** overriding Claude's "measurement is broken" nudge is a live proof-of-concept of the thesis — the argument enacts its own prescription. This is rhetorically strong and epistemically honest. - **Causal claim (Part 5):** "offload → skill rusts → cheaper to replace" is a plausible causal chain; strengthen with evidence on skill-decay rates rather than assertion. #### Fallacies - **False equivalence / unit mismatch — "The math doesn't math" (Part 3, deployment):** 29% of *companies* seeing significant ROI and *individuals* claiming 5x productivity are different units at different levels of analysis; they can both be true without contradiction. Fix: state them as two separate observations (org-level ROI is rare; self-reported individual gains are high and possibly inflated) rather than implying an arithmetic contradiction. - **Hasty generalization / unrepresentative sample (Part 3, throughout):** Reddit verbatims from self-selecting, venting communities are treated as representative of practitioners generally. Fix: qualify scope ("among vocal Reddit participants…") or triangulate with survey data. - **Appeal to nature (Part 2, "Executive function is one of the things that defines our sentience"):** using sentience/animal cognition to imply we *ought* not cede executive function slides from *is* to *ought*. Fix: argue the practical harm (accountability, deskilling) directly, which the piece already does — the sentience claim is decorative, not load-bearing. - **Unsupported statistic (Part 3, sycophancy):** "~90% accuracy" purchase-intent emulation is asserted without citation. Fix: link the academic source or hedge the figure. - **Suggestive attribution / possible speculation (Part 3, budget):** "this is by design to create habits" attributes intent to vendors on one commenter's say-so. Fix: mark as hypothesis and note it's a design *effect* regardless of intent. #### Tone of voice - Conversational, candid, lightly polemical, self-aware. Examples: "foaming at the mouth on Reddit," "meat-based interface to an LLM," "panic city." Warm second-person inclusivity ("I appreciate that YOU are") balances the critique; confident but not contemptuous. ### 🛡️ DARVO Analysis - Not applicable to this source. Reason: DARVO (Deny, Attack, Reverse Victim and Offender) describes a defensive response to an accusation; this is a first-person analytical essay with no accuser, no accused party defending itself, and no victim/offender reversal. Penn implicates "we" (including himself) rather than deflecting blame, so none of the three DARVO moves are present. ## 🎯 Conclusion - **Bottom line:** The biggest problem in AI isn't cost, sustainability, or ethics — it's abdicated executive function. When the machine decides, no one is accountable, skills atrophy, and pride and value drain out of the work. The cure is to lead the tasks that matter and make the tool educate you. - **Practical next steps for readers:** (1) Apply the two-question gate — does AI make the output better *and* make me better — to your next real task. (2) Name a human decision-of-record on AI-assisted deliverables. (3) For anything in your expertise, dictate your own read first, then have AI critique and quiz you. (4) Outside your expertise, run agentic deep research, build a quality checklist, and self-teach the domain. (5) Watch your defaults — model choice, skip-permissions, "AI deployed" claims — and decide them on purpose. ## 🔄 Refactor Appendix — Almost Timely News: The Biggest Problem with AI Today > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults: 3 Literature, 1 Permanent. Resonance filter: if you can't > articulate in one line why a claim matters, drop it. > > Konik claim-header rules: present tense, declarative, ≤10 words when > possible, searchable, unique, mirrors annotation focus if present. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Atoms (H3) #### [H1] AI's real problem is abdicated executive function > [!quote] ID:rw_penn20260705a > we've abdicated far too much of our executive function to AI. We've surrendered our thinking. **Context & use**: - Part 1 (intro / thesis statement). - Reframes cost/sustainability/ethics as symptoms of a cognition-and-accountability failure; the article's spine. **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H2] Lead first, then force the machine to educate you > [!quote] ID:rw_penn20260705b > for every task that matters, always start with something you lead, and force the machines to educate you. **Context & use**: - Part 4 (The Antidote). - The prescriptive core: use AI to grow expertise, not substitute for it — the operational antidote to deskilling. **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H3] Two-question gate: better output and better you > [!quote] ID:rw_penn20260705c > 1. Does the use of AI make the output better? 2. Does the use of AI make me better? **Context & use**: - Part 5 (Wrapping Up). - A portable decision test; must be yes to both or the tool is being used badly. Strong editorial/review checkpoint. **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Outsourcing judgment removes the accountable decision-maker **Claim** (one sentence, Courtney's voice): When you delegate executive function to AI, no human owns the decision, so accountability, skill, and professional value quietly erode. **Evidence pointer**: [[#H1]] > [!quote] > What's strange is nobody decided this. There was no meeting where we discussed this. **Scope & conditions**: Holds for tasks that matter — client deliverables, analysis, decisions with stakeholders. Breaks for low-stakes, fully-commoditized tasks where understanding the output carries no downstream cost. **Links** (≥1 with annotation): - [[The Problem with Streaming and the Case for Physical Media — Tom Rizzuto TED]] — because extends (Rizzuto: don't outsource ownership of your media; Penn: don't outsource ownership of your thinking — same agency-and-ownership axis) - [[Metacognition — thinking about thinking]] — because applies-in professional AI workflows - [[Deskilling under automation]] — because contrasts on whether the cause is the tool or the ceded decision <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Replicate Penn's AI-detector false-positive test (he got ~1-in-7 on human text) at larger N. — source Part 3 #fleeting #inbox - [F2] Draft a "decision-of-record" review-gate template for AI-assisted deliverables. — #articleseed #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Add the two-question gate (better output AND better me) to my content review checklist — editorial ops #things - [ ] Watch Tom Rizzuto's streaming/physical-media TED talk and note the ownership-analogy parallels for the cross-link — reading #things ``` Processed: Almost Timely News — The Biggest Problem with AI Today (2026-07-05) - Literature atoms: 3 / 5 cap - Permanent candidates: 1 / 3 cap - Fleeting items: 2 - Things 3 actions: 2 / 2 cap - Konik tags used: [#xref, #aphorisms, #articleseed] - Triage callouts populated: [Cross Reference, Article Ideas] - Quality gates: [verbatim quotes ✓, atomicity ✓, link annotations ✓, YAML hygiene ✓, Konik triage ✓, view-highlight markers ✓] - Override conditions triggered: none ``` <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Almost Timely News 🗞️ the Biggest Problem With AI Today.md|Almost Timely News: 🗞️ the Biggest Problem With AI Today by Christopher S. Penn]] ## Highlights - Or more correctly, we've abdicated far too much of our executive function to AI. We've surrendered our thinking. ([View Highlight](https://read.readwise.io/read/01kwsmdtt24tmmxx8gwk2vs9at)) - Note: The thesis: cost/sustainability/ethics are symptoms; the root problem is abdicated executive function. - Tags: atom/literature/h1 - As a reminder, I bucket executive function into four categories that I call PODS: Plan – you think about achieving something in the future and make a plan to get there from here; Organize – you take what you have and try to make sense of it; Decide – you take what you have and make decisions about it; Solve – you solve the problems you have. ([View Highlight](https://read.readwise.io/read/01kwsme0w4mda6y1v177acy2aq)) - Note: PODS — the operational handle for executive function (Plan, Organize, Decide, Solve). - Tags: atom/zettel/z1 - AI is hollowing out companies: I really liked this quote from the agency owners subreddit: "What's strange is nobody decided this. There was no meeting where we discussed this. We automated one annoying task, then another, and one day the job had hollowed out from the inside." This erosion of tasks is all about a lack of cognition, a lack of reflection, a lack of a plan. No one's making decisions – just leaving it up to the machines, a bit more each day. ([View Highlight](https://read.readwise.io/read/01kwsmdygdvjbxdxqw5480a5pt)) - Note: "Nobody decided this" — the accountability vacuum that erosion-by-automation creates. - Tags: atom/zettel/z1 - So my suggested antidote is this: for every task that matters, always start with something you lead, and force the machines to educate you. ([View Highlight](https://read.readwise.io/read/01kwsmdvvqkxrgg9ae6g1qb0ev)) - Note: The prescriptive core: use AI to grow expertise, not substitute for it — the operational antidote to deskilling. - Tags: atom/literature/h2 - We can boil it all down to a simple set of questions: 1. Does the use of AI make the output better? 2. Does the use of AI make me better? If the answer isn't yes to BOTH, then you're not using it well. ([View Highlight](https://read.readwise.io/read/01kwsmdwemd3713462qqv2aabt)) - Note: The two-question gate — a portable decision test / editorial checkpoint. Must be yes to both. - Tags: atom/literature/h3 ## Links - [[AI-Assisted PKM Replaces Synthesis Work With the Appearance of Synthesis Work]] — existing Permanent captures the cost of handing executive and synthesis work to the tool. - [[AI Literacy Is Relational Not Just Technical]] — applies-in: the two-question gate evaluates what the tool does to the user, not only what it produces.