AI Fatigue Is Coming  🔗 **Source:** [AI Fatigue Is Coming](https://davidspinks.substack.com/p/ai-fatigue-is-coming) > [!note]+ Document Notes > tags: > - ai-fatigue > - ai-adoption > - burnout > - community-building > - human-connection > author: David Spinks > source: https://davidspinks.substack.com/p/ai-fatigue-is-coming > reference: https://davidspinks.substack.com/p/ai-fatigue-is-coming > date created: 2026-09-15 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > <!-- deep-read:sections start --> ## 🏷️ Title **AI Fatigue is Coming** — David Spinks, personal Substack, ~860 words. A founding figure in community building (CMX) argues that heavy daily AI use produces a specific, novel exhaustion, and describes rolling his own usage back. ## 🧭 Table of Contents Title · Contents · References (in-page) · Summary · Main Content (Notes) · References · Paraphrase · Newsletter · Action Items · Hypotheses · Methodology · Questions · Future Research · Implications · Learning Styles · Accessibility · DEIB · Logical Validity · DARVO · Conclusion · Refactor Appendix ## 🔗 References (in-page) Exactly two outbound links; everything else is unlinked assertion. - Library of Congress item 2020715038 — https://www.loc.gov/item/2020715038/ (header image credit) - DavidSpinks.com — http://davidspinks.com (the author's coaching practice, closing CTA) No studies, surveys, or third-party reporting are cited. The only named tool is Claude, mentioned once as tracking his to-do list. ## 📝 Summary **What happened.** After months of daily AI use Spinks reports brain fog, body tiredness, and disembodiment, and names four mechanisms: *amplified multitasking* (starting is cheap, so he starts too much; he cites his ADHD), a *verification tax* (it errs, ignores his rules, needs extra review rounds), *atrophy and agency drift* (planning outsourced until the tool starts making business decisions for him), and *empty social calories* — the genuinely new part: hours his mind reads as social that carry no social nutrition. He concedes the first three are not AI-specific, reports a partial rollback, doubts he has cut enough, and predicts fatigue is coming for everyone. **Why it matters.** A trusted community voice publishing a defection narrative at peak adoption pressure. If power users hit fatigue first, the curve reaches communities before analyst reports. **Stakeholders.** Community and DevRel staff told to ship AI-assisted work faster while owning a room's emotional weather; contributors under usage mandates; maintainers absorbing AI-generated patches. **Risks.** n-of-1 generalized to everyone; usable as anti-AI ammunition or as grounds to dismiss worker complaints as adjustment pains. An unfacilitated "AI fatigue" thread becomes a venting spiral. **Bottom line.** Trust the symptom inventory, not the forecast. ## 📖 Main Content The structure: a hook asking readers to check their own body, four diagnostic passes, one concession, one novelty claim, a remedy list, and a closing question that turns a personal essay into a community prompt. The strongest move is the concession — admitting pre-existing weaknesses removes the "that's a you problem" rebuttal and narrows the novelty claim to the social mechanism alone. The second is that the remedy is asymmetric: he evicted AI from creative work and decisions and kept it where it performs. That is a portfolio decision, not a moral stance. The weakest third is the ending, where a bounded account becomes a universal forecast with no bridge. ### 🗒️ Notes - **The verification tax is the measurable part.** Time-to-accept, rework rounds, share discarded — log it and argue dosage with data instead of vibes. - **"Empty social calories" is the phrase that will travel.** A compact frame for parasocial substitution that names what community work actually sells. - **Agency drift is the quietest risk.** Not "it does my work" but "it's started to make decisions for me" — governance, not productivity. AI-assisted triage quietly becomes the roadmap. - **Mode collapse in feedback.** Advice that "locks onto the same feedback norms" is a warning for AI in mentorship, contributor review, and onboarding, where specificity is the whole value. ## 📚 References **Primary (present in the source):** - David Spinks, "AI Fatigue is Coming" — https://davidspinks.substack.com/p/ai-fatigue-is-coming - Library of Congress item 2020715038 — https://www.loc.gov/item/2020715038/ (the only external institution linked) - DavidSpinks.com — http://davidspinks.com (the author's stake: he sells human coaching) No claim rests on outside work, so there is no literature to chase on the essay's behalf. Rather than manufacture citations the author did not make, this section stops here. ## 🗣️ Paraphrase *Paragraph by paragraph, in the analyst's words.* 1. Do you also feel foggy, disconnected, and body-tired after months of daily AI use? 2. He has noticed it in himself and now sees the sources. 3. **Mechanism one.** AI runs his planning, yet he juggles too much because starting is cheap — ADHD running hot. 4. **Mechanism two.** It errs, ignores his rules, needs repeated review; like someone who half listens. 5. **Mechanism three.** Savings are real, but long outputs and constant management consume them; he has left family time to unblock agents. 6. **Mechanism four.** The to-do list costs the feel of his day, email the writing muscle, planning has let the tool shape decisions. 7. **The concession.** Multitasking, time management, and delegation were already weak points. 8. **The novelty claim.** Hours go to a robot, not a human with feelings and a shared social web. Empty social calories. 9. **The symptom.** Disconnection and disembodiment, like a social media hangover but stranger. 10. **Novelty decay.** Once the wow factor faded, awe became impatience. 11. **Ambient load.** Constant management at both ends of the day, promises unmet. 12. **The rollback.** No creative work, some processes manual again, no delegated decisions — and doubt it is enough. 13. **Whittling.** He adopted wholesale and is trimming to the few things it does well. 14. **Forecast and invitation.** He predicts broad fatigue soon and asks whether it is happening to you. ## 📰 Newsletter **Empty social calories, and what they cost a WordPress community** David Spinks — who professionalized community building with CMX — just published ["AI Fatigue is Coming"](https://davidspinks.substack.com/p/ai-fatigue-is-coming). It is not a hot take, it is an inventory: foggy, tired, re-reviewing output that ignores the rules he set, sneaking away from family to approve agent requests. Every contributor who has "just quickly triaged one more ticket" during dinner knows that tug. So, a permission slip. If your Meetup's next three topics are all AI, swap one; run a session where people bring the workflow they *stopped* using. In [Training](https://learn.wordpress.org/), resist opening every module with a chatbot. In Support and Docs, if AI drafts cost more review rounds than they save, write that number down instead of arguing vibes in [Slack](https://make.wordpress.org/chat/). For Core AI and [Performance](https://make.wordpress.org/performance/), the sharpest warning is drift: keep the human on the decision, the machine on the draft. Bring snacks. Touch grass. ## ✅ Action Items ### DevRel - Add one pulse item: *"In the last month, has AI-assisted work left you more tired, less tired, or about the same?"* Local data before the 2026 calendar is set. - Log the verification tax (rework rounds, time-to-accept, percent discarded) beside time-saved claims. - Protect one flagship human-to-human program per quarter — office hours, pair debugging, mentor matching — defended on connection and retention, not reach. ### OSPO - Guidance on agentic tooling and off-hours interrupts: approving agent actions after hours is not an expectation. - Put the drafting-versus-deciding boundary in policy — AI may draft, summarize, and triage; humans own roadmap and prioritization. - Track AI-assisted contribution review burden so the cost lands in the same ledger as the throughput gain. ### Community - Pilot a "what I un-adopted" session at one Meetup or contributor day, facilitated to stay constructive. - Audit AI saturation in the event calendar and rebalance; novelty decay is coming for that programming. - Ensure first-time contributors meet a human early — mode-collapsed feedback is the opposite of what they need. ### WordPress subteams - **Core / Core Program Team / Core AI** — Name a human decision owner for anything shaping release scope; treat fatigue as a design input (shorter output, surfaced uncertainty, rules that stay respected). - **Core Performance / Tide / Meta** — Measure review overhead on AI-assisted patches as rigorously as benchmarks; automate the audit, keep interpretation human. - **Training / Documentation / Polyglots** — Do not front-load AI into every path; specificity is the product, and machine assistance raises the same review-burden question. - **Support / Hosting / Themes / Plugins** — Review queues are where the tax shows first; never let generated content displace the human escalation path. - **Community / Marketing / TV / Design / Accessibility / Mobile** — Resist wow-factor framing; long unstructured output is an accessibility problem, and approval pings are the new pager. - **The WPOCC (The WP Open Community Collective)** — Well placed to convene a cross-project conversation on sustainable AI practice and contributor wellbeing. - **FAIR** — Federated governance needs explicit norms on AI-assisted decision-making so drift does not enter through automation. ## 🧪 Hypotheses 1. **Verification tax.** Logged rework rounds put net savings below claimed savings. *Test:* instrument one workflow for six weeks, assisted versus unassisted, on time-to-accept and discard rate. 2. **Power users first.** Fatigue reports rise with usage intensity. *Test:* cross-tabulate usage frequency against a fatigue item in a community pulse. 3. **Social substitution.** High-AI, low-human days predict greater depletion than equal-workload days with more human contact. *Test:* two-week diary study. 4. **Mode collapse.** AI feedback converges on a narrow recommendation set regardless of actual weaknesses. *Test:* blind-score fifty AI reviews against matched human reviews. ## 📊 Methodology An opinion essay; its evidence base should be stated plainly rather than dressed up. **What it rests on:** personal introspection over a period given only as "months in a row"; autobiographical anecdotes (planning, email, coaching supervision, leaving family time to unblock agents); analogy to the social-media hangover; and extrapolation from one case to a population, which the author flags as a guess. **What is absent:** no survey, interviews, usage logs, time tracking, comparison group, or defined period, and no citation to work on cognitive load or automation complacency. The single external link is an image credit. **Weight:** high as hypothesis generation and sentiment signal; low as evidence of prevalence. ## ❓ Questions - At what daily intensity, and over how long, did symptoms develop? "Months in a row" is the only dosage detail. - Which mechanism dominates? If the verification tax vanished, would the disembodiment remain? - Is this AI-specific or screen-specific? The social-media analogy raises the confound and leaves it open. - Does it matter that the author sells human coaching, and how much should that discount the argument? - What would count as disconfirmation? No condition is offered. ## 🔮 Future Research - Longitudinal usage-and-wellbeing panel among high-intensity adopters, tracking depletion against measured rework. - Controlled comparison of AI-mediated versus human-mediated onboarding feedback, measured on 30- and 90-day contributor retention. - Instrumented study of the verification tax by task type, to find which categories net out positive. - Organizational study of agency drift: how often AI-generated prioritization becomes the accepted plan with no explicit human decision. ## 💡 Implications **Practice.** Human-to-human formats are the differentiated product, not legacy overhead; facilitation skill gains value because the sessions worth running are the ones a chatbot cannot hold. **Policy.** Norms on off-hours agent approvals, and a named line between drafting and deciding. "AI may propose, humans dispose" answers the essay's most serious complaint in one sentence. **Security.** Automation complacency is a security posture: a fatigued reviewer who has stopped reading long outputs closely approves the bad patch or the over-scoped permission. **Education.** Preserve productive struggle: if the learner never crafts the thoughtful message, the muscle never forms, and assessment that measures only output is measuring the tool. **Funding.** Human-centered programming needs justification on connection, retention, and trust — returns that reach dashboards more slowly than content volume. ## 🧠 Learning Styles Assessment - **Visual** — Weak. One decorative header image, no diagrams; a four-mechanism diagram would carry the argument better than any paragraph. - **Aural** — Strong. Conversational and rhythmic; it would work nearly unchanged as a podcast monologue. - **Verbal** — Strong, and the primary channel. Coined phrases do the persuasive work. - **Kinesthetic** — Moderate to strong. Symptoms are bodily, and the remedy list is copyable this week. - **Logical** — Weak to moderate. Mechanisms are cleanly separated, but there is no data and the leap from self to everyone is unsupported. - **Social** — Strong. Opens and closes with questions to the reader; built to start a conversation. - **Solitary** — Strong. Short, self-contained, structured as a self-audit prompt. - **Naturalistic** — Light. Nature appears as antidote ("touch grass, tickle some trees") but is a gesture, not a thread. - **Reading level** — Roughly US grade 8–9; broadly accessible, though one instance of profanity may need softening for some workplace contexts. ## 📊 Accessibility Assessment Assessed against WCAG 2.2 AA as captured. 1. **Text alternatives (1.1.1)** — Fails: both images carry empty alt text. *Fix:* descriptive alt on the header, decorative marking on the banners. 2. **Info and relationships (1.3.1, 2.4.6)** — Weak. A flat run of paragraphs with one trailing heading; nothing to navigate the four mechanisms by. *Fix:* headings matching the argument. 3. **Contrast (1.4.3, 1.4.11)** — Not assessable from a text capture; Substack defaults generally meet AA for body text. *Fix:* verify the live page. 4. **Keyboard and focus (2.1.1, 2.4.7, 2.4.11 Focus Not Obscured)** — Not assessable. Subscribe overlays are a known focus-obstruction risk. *Fix:* keyboard-only pass with the modal open. 5. **Link purpose (2.4.4)** — Mixed. The two text links are clear; the image-wrapped links are unlabeled. *Fix:* label or suppress them. 6. **Readable language (3.1.1, 3.1.5)** — Good. Plain English at grade 8–9; confirm the page language attribute. Idioms like "touch grass" may not translate. 7. **Name/role/value and status messages (4.1.2, 4.1.3)** — Not assessable. Subscribe widgets and like counters are the usual offenders. *Fix:* verify dynamic confirmations are announced. ## 🌍 DEIB Assessment - **Cultural breadth** — Narrow. A US-based, self-employed worker who controls his own tooling, schedule, and adoption level. Workers under a usage mandate, and those where these tools are gated or costly, have a different relationship to this fatigue. - **Inclusive language** — Generally good; first-person framing avoids prescribing. "Touch grass" assumes outdoor access and mobility. - **Neurodiversity support** — Double-edged. Naming his ADHD and describing AI amplifying it is honest and rare, but there is no counterweight from people for whom these tools reduce executive-function load. - **Representation** — Limited by design. One voice, with no gesture at whose experience might differ. - **Psychological safety** — High, and the essay's clearest strength. Admitting fatigue and over-adoption under adoption pressure gives readers permission to say the same; that modeling is the transferable asset. ## ✅ Logical Validity **Where the reasoning is strong.** The four mechanisms are distinct and non-overlapping. The concession isolates the variable and narrows the novelty claim to the social mechanism, pre-empting the strongest objection. Claims are properly hedged. **Fallacies.** 1. **Hasty generalization (n = 1 → everyone).** *Located:* the closing prediction, "I can't imagine I'm alone… AI fatigue is coming for everyone sooner rather than later." *Why:* one self-observed case cannot support a population forecast, and "I can't imagine" is incredulity standing in for evidence. *Fix:* scope it — "among daily high-intensity users I expect fatigue reports to rise this year" — and pair it with a reader poll. 2. **Sweeping generalization.** *Located:* "And then AI always seems to get things wrong and require extra rounds of reviews." *Why:* "always" overstates a claim the essay itself contradicts by granting that some hours-long tasks now take minutes. *Fix:* bound it — "roughly half need another pass" — which strengthens the complaint. **Tone of voice.** Confessional, conversational, deliberately unpolished, persuading by relatability rather than evidence and honest about doing so — the register practitioners use to open a hard conversation, and also what lets the forecast slide past unchallenged. ## 🛡️ DARVO Analysis - **Deny** — Not present. The essay grants that AI genuinely saves time and that his own struggles with multitasking, time management, and delegation contribute. - **Attack** — Minimal and impersonal. Criticism targets a tool's behavior; the sharpest lines anthropomorphize software rather than disparage vendors, adopters, or colleagues. - **Reverse victim and offender** — Not present in the manipulative sense. He names himself as the adopter who went in wholesale and assigns himself the remedy, never claiming victimhood. The nearest edge is rhetorical — the tool "slowly" drawing him into reliance — but he reclaims responsibility in the next paragraph. **Verdict:** no DARVO pattern. A self-implicating account that can be engaged on its claims without navigating defensive rhetoric. ## 🎯 Conclusion **Bottom line.** A careful symptom inventory attached to a weak forecast; grade the two separately. The mechanisms are specific enough to instrument, and "empty social calories" is the one that most concerns anyone whose job is building rooms where humans connect. The prediction is a hunch and should not be cited as evidence of a trend. **Next steps.** (1) Add a fatigue item to the next pulse. (2) Log the verification tax beside time-saved claims. (3) Write the drafting-versus-deciding boundary down before drift decides for you. (4) Pilot "what I un-adopted" at one Meetup. ## 🧱 Refactor Appendix ### Literature Split Pad #### H1 — Empty social calories > [!quote] But what makes the brain fog and exhaustion of using AI different is that all of this time, hours every day, is spent talking to a robot. I’m not problem-solving with another warm-blooded human with feelings, depth, and a shared social web. I’m talking to a semi-intelligent wall. I’m tricking my mind into thinking its getting social interaction but it’s not. It’s empty social calories. It’s loneliness fuel. The only claim to novelty, and the passage with the most direct bearing on community practice. `#atom/literature/h1` #### H2 — The verification tax and the always-on tug > [!quote] AI is supposed to save me time and energy, and it does. There are some tasks that would have taken me hours that I can do in just minutes. Yet, I find myself spending loads of time and energy reading through its lengthy responses and managing it throughout the day. I’m definitely not less busy. And the worst part is that I found myself sneaking off during family time to approve requests and keep my agents moving. Real savings conceded, and the management overhead that eats them. `#atom/literature/h2` #### H3 — Whittling back from wholesale adoption > [!quote] I feel like I adopted it wholesale, and now I’m whittling it back to the core things it can do well, that don’t exhaust me, and keep the rest in the human realm. Part of me wants to just ditch the whole thing. Do I really need this? Rollback as portfolio decision, not moral stance — the transferable pattern for team policy. `#atom/literature/h3` ### Atomic Split Pad #### Z1 — Parasocial substitution depletes the budget community work refills If machine hours register as social contact without delivering social nutrition, AI-heavy work draws down the connection budget community programming exists to restore. Human formats therefore become more valuable as AI use rises, and should be defended on connection and retention rather than reach. `#atom/zettel/z1` #### Z2 — Agency drift is a governance failure wearing productivity clothes "It's started to make decisions for me" differs categorically from "it does my work faster." Drift happens without any decision to delegate authority, which is why it needs a written boundary rather than good intentions. `#atom/zettel/z2` ### Fleeting Queue - Does fatigue follow usage intensity, or usage *mandate*? A self-employed coach who chose wholesale adoption differs from an employee with a target. - The security angle is unwritten: fatigued reviewers approving long outputs they have stopped reading closely is a supply-chain risk, not just a wellbeing one. `#atom/fleeting/f1` <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/AI Fatigue Is Coming.md|AI Fatigue Is Coming by David Spinks]] ## Highlights - And then AI always seems to get things wrong and require extra rounds of reviews. ([View Highlight](https://read.readwise.io/read/01m2jwgx48hrewfnmpy3zjykze)) - Note: Sweeping generalization: "always" contradicts his own concession that some hours-long tasks now take minutes. - Tags: fallacy/sweeping-generalization - AI is supposed to save me time and energy, and it does. There are some tasks that would have taken me hours that I can do in just minutes. Yet, I find myself spending loads of time and energy reading through its lengthy responses and managing it throughout the day. I’m definitely not less busy. And the worst part is that I found myself sneaking off during family time to approve requests and keep my agents moving. ([View Highlight](https://read.readwise.io/read/01m2jwgqbvqrq7rpm0edvyycef)) - Note: The verification tax is the one mechanism a team can instrument — log rework rounds next to time-saved claims. - Tags: action/h2, atom/literature/h2 - But what makes the brain fog and exhaustion of using AI different is that all of this time, hours every day, is spent talking to a robot. I’m not problem-solving with another warm-blooded human with feelings, depth, and a shared social web. I’m talking to a semi-intelligent wall. I’m tricking my mind into thinking its getting social interaction but it’s not. It’s empty social calories. It’s loneliness fuel. ([View Highlight](https://read.readwise.io/read/01m2jwgmwn56rzqpaz5re397dc)) - Note: The essay's only novelty claim, and the one that bears directly on community practice: machine interaction registers as social contact but delivers no nutrition. - Tags: atom/zettel/z1, atom/literature/h1 - I feel like I adopted it wholesale, and now I’m whittling it back to the core things it can do well, that don’t exhaust me, and keep the rest in the human realm. Part of me wants to just ditch the whole thing. Do I really need this? ([View Highlight](https://read.readwise.io/read/01m2jwgse1xtxqxjgpaq4gvcwg)) - Note: Rollback as a portfolio decision rather than a moral stance — the transferable template for a team AI-use policy. - Tags: atom/literature/h3, action/h3 - I can’t imagine I’m alone. This is probably happening for a lot of people if it’s happening for me. My guess is AI fatigue is coming for everyone sooner rather than later. ([View Highlight](https://read.readwise.io/read/01m2jwgvdfnve95ctmr2y0mmg0)) - Note: Hasty generalization: one self-selected case plus personal incredulity standing in for evidence of a population trend. - Tags: fallacy/hasty-generalization, atom/fleeting/f1