The True Cost of AI Coding - CourtneyR.dev's 2nd Brain

The True Cost of AI Coding ![rw-book-cover](https://i.ytimg.com/vi/iPUn1Fnfn0k/sddefault.jpg) 🔗 **Source:** [The True Cost of AI Coding](https://www.youtube.com/watch?v=iPUn1Fnfn0k) > [!note]+ Document Notes > tags: [ai-coding, developer-burnout, mental-health, ai-productivity, workplace-wellbeing] > author: Scott Tolinski > source: Syntax.fm (YouTube) > reference: https://www.youtube.com/watch?v=iPUn1Fnfn0k > date created: 2026-08-14 00:00 > > > [!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 Do not quote or attribute any other line to the named author. > > # 🏷️ Title: The True Cost of AI Coding >  ## 🧭 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 - [Take the survey](https://forms.gle/TvtjgtU7UPnK7DC9A) - [Charts / data site](https://ai-health.syntax.fm) - [Dr. Courtney Tolinski, PhD, LP, NCSP — Phases podcast](https://phases.fm) - [Adam Elmore](https://x.com/adamdotdev) - [Mark Erikson](https://bsky.app/profile/acemarke.dev) / [blog](https://blog.isquaredsoftware.com) - [Miranda Heath](https://bsky.app/profile/mirandaheath.website) - [Aaron Francis](https://x.com/aarondfrancis) / [site](https://aaronfrancis.com/) - [Stanford AI Index Report 2026 (PDF)](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) - ["The brain side of human-AI interactions in the long-term: the '3R principle'" — Nature](https://www.nature.com/articles/s44387-025-00063-1) - [ADAA.org](https://adaa.org) — mental health crisis resources cited in the video (988 in the US) ## 📝 Summary - **What happened.** Scott Tolinski (Syntax.fm) surveyed nearly 1,300 developers and interviewed a psychologist, a burnout survivor, a veteran maintainer, a burnout researcher, and a self-employed developer with no negative side effects, to investigate how heavy AI-assisted ("prompt") coding affects developers' mental health — sleep, burnout, skill perception, enjoyment, and identity. - **Why it matters.** This is original survey data (not just anecdote or vendor marketing) on a real and under-discussed cost of AI coding adoption: 46% of respondents regularly prompt past their intended stopping point, 58% of that group report sleep disruption, 59% feel their coding skills are diminishing, and 54% report less enjoyment — with the video explicitly noting the skill/enjoyment findings come from largely the same people. - **Key stakeholders.** Individual developers (subject of the survey and interviews); employers/managers (implicated in the "pressure to produce more" findings and the structural argument at the end); AI tool vendors (implicit beneficiaries of increased usage/pressure); mental health professionals (framing the addictive mechanism and coping strategies). - **Risks and tradeoffs.** The survey is self-selected (people who wanted to talk about this), which the video explicitly and repeatedly flags — a real methodological limitation that could get lost if these percentages are cited elsewhere without that caveat. The video opens with an extreme, viral anecdote (a developer hospitalized after an all-night prompting binge) before introducing the more measured survey data. - **Bottom line.** Heavy AI-coding usage correlates with worse sleep, more perceived pressure, and (for many) diminished skill/enjoyment — via a "variable reward" dopamine mechanism a psychologist compares to slot machines — but the video's own strongest point is structural, not personal: individual coping strategies (time limits, boundaries) only work if the surrounding workplace culture allows people to actually stop. ## 📖 Main Content - The video opens with a viral tweet about a developer ("Rob") hospitalized after staying up all night trying to burn through AI tokens before a model was rumored to be shut down — used as a hook before pivoting to systematic survey data (nearly 1,300 developer responses) and expert interviews. - **Burnout case study:** Adam Elmore (17-year developer, open-source maintainer) describes a major burnout preceded by intense AI usage — sleeplessness, compulsive prompting, loss of excitement to work, eventually needing to step back entirely. - **The addiction mechanism:** Dr. Courtney Tolinski (licensed psychologist) explains burnout symptoms (sleep loss, anxiety, social withdrawal, somatization) and frames compulsive prompting as a variable-reward system analogous to a slot machine — intermittent "perfect solution" rewards drive dopamine-fueled repeated engagement ("just one more prompt"). - **Sleep and stopping-point data:** 46% of respondents regularly prompt past their intended stopping point; of that group, 58% report sleep changes (vs. 17% among those who stop on time) — described as the survey's strongest relationship, with an explicit correlation-not-causation caveat. - **Pressure to produce more:** citing the 2026 AI Index report's finding that GitHub Copilot users completed 26% more pull requests, the survey found 65% of respondents feel more pressure to produce more often or daily. Mark Erikson (Redux maintainer) frames this as a capitalism-driven "if we can go faster, we must go faster" dynamic, extending naturally toward running many parallel AI agents. - **Skill, enjoyment, and identity:** 59% feel their coding skills are diminishing; 54% report less enjoyment than before — largely the same respondents for both. Adam Elmore describes rebuilding a deliberately slower workflow that keeps him "driving the work" and reviewing code himself, tying this to a cited research framing: "AI gives you a result, but a result isn't an answer yet. It becomes one when a person evaluates it and takes responsibility for it." Miranda Heath (burnout researcher) ties skill loss to neuroplasticity — unused skills atrophy. - **What separates healthy from unhealthy use:** Aaron Francis (Laravel VP of marketing/community, self-employed) reports no negative mental health effects, attributing this to self-imposed boundaries (time limits, offline hours, family/church rhythms, no prompting apps on his phone) — but the video explicitly notes these boundaries are easier to hold because he has no manager or quota pressure. - **Agent-count data:** most respondents run 1-2 AI agents simultaneously; running more agents correlates with both higher enjoyment and more sleep disturbance, and (counterintuitively) with feeling less skill diminishment — the video explicitly flags this last finding as too data-thin (few respondents running 4-5 agents) to draw conclusions from. - **Structural closing argument:** citing 2026 AI Index employment data showing a ~20% decline in software developer employment for ages 22-25 since 2022 (and a larger relative decline in AI-exposed roles specifically) — while carefully noting unemployment rose across all job categories in the same window, so this is not proof AI caused the decline — the video argues young/less-secure developers face more pressure to override healthy boundaries, and concludes that individual discipline can't fix a problem that is partly structural (workplace expectations, job security). ### 🗒️ Notes #### ⏱️ [00:00](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=0) Intro - Cold open: viral tweet about a developer hospitalized after an all-night AI token-burning session. #### ⏱️ [00:20](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=20) Rob in the hospital from prompting - Context on the viral incident and mixed public reaction (support vs. confusion). #### ⏱️ [01:08](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=68) Disclaimer - Framing: survey of ~1,300 developers, reflects survey respondents, not the whole developer population. #### ⏱️ [01:11](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=71) Chapter 1: Crash & Burnout - Adam Elmore's burnout story; Dr. Tolinski on burnout symptoms and somatization. #### ⏱️ [05:29](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=329) Chapter 2: Why is prompting hard to put down? - 46%/58%/17% stopping-point and sleep data; variable-reward/slot-machine dopamine mechanism. #### ⏱️ [07:39](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=459) Chapter 3: If AI saves time, why is there more pressure? - 26% more PRs (AI Index), 65% feel more pressure; Mark Erikson on the "must go faster" dynamic and parallel agents. #### ⏱️ [10:23](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=623) Chapter 4: What happens to enjoyment, skill, and identity? - 59% diminishing skills, 54% less enjoyment; Adam Elmore's deliberate slower workflow; neuroplasticity framing. #### ⏱️ [14:00](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=840) Chapter 5: What separates energizing use from unhealthy use? - Aaron Francis's boundaries; agent-count vs. enjoyment/sleep/skill data (with explicit thin-data caveat). #### ⏱️ [17:52](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=1072) Chapter 6: What should change? - Dr. Tolinski's concrete coping strategies; Aaron Francis's specific boundary rules; structural/employment-data closing argument. #### ⏱️ [24:21](https://www.youtube.com/watch?v=iPUn1Fnfn0k&t=1461) Support - Survey still open; call for viewer input on what's working/not working for them. ## 📚 References - [Stanford AI Index Report 2026 (PDF)](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) — primary source, PR/employment data - ["The brain side of human-AI interactions in the long-term: the '3R principle'" — Nature](https://www.nature.com/articles/s44387-025-00063-1) — primary source, cited for the "result isn't an answer until a person takes responsibility for it" framing - [Survey (still open)](https://forms.gle/TvtjgtU7UPnK7DC9A) — primary source, self-run by Syntax.fm ## 🗣️ Paraphrase - Opening: a viral anecdote about a hospitalized developer sets up a broader investigation into AI coding's mental health costs, grounded in a nearly-1,300-respondent survey and expert interviews. - Burnout section: intense AI usage preceded a serious personal burnout for one interviewee; a psychologist explains the symptom profile and frames compulsive prompting as behaviorally similar to slot-machine variable-reward addiction. - Pull section: nearly half of respondents regularly overshoot their stopping point, and that group reports far more sleep disruption than those who stop on time — the strongest relationship found, though explicitly not proven causal. - Pressure section: even though AI measurably increases output (26% more PRs per external data), two-thirds of respondents feel more pressure to produce, which one interviewee ties to a broader "go faster because we can" industry dynamic. - Identity section: skill and enjoyment both declined for a majority of the same respondents; interviewees discuss maintaining a slower, more deliberate workflow and the neuroplasticity risk of skills atrophying from disuse. - Boundaries section: a self-employed interviewee describes personal limits that prevent burnout, but the video notes these boundaries are more available to him than to developers under managerial/quota pressure. - Closing section: employment data (carefully caveated as non-causal) is used to argue that younger, less-secure developers face outsized pressure to override healthy limits, and that individual discipline alone can't solve what is partly a structural/workplace problem. ## 📰 Newsletter - **Worth a watch if you use AI coding tools daily.** Syntax.fm's Scott Tolinski surveyed nearly 1,300 developers on how AI-assisted coding is affecting their mental health, and the numbers are sobering: 46% regularly prompt past their intended stopping point, and that group reports sleep disruption at more than 3x the rate of people who stop on time. 59% feel their coding skills are slipping; 54% enjoy coding less than before. A psychologist frames compulsive prompting as a slot-machine-style variable-reward loop. The most useful part might be the closing argument, though: individual boundaries (time limits, no-phone rules, offline hours) only work if your workplace culture actually lets you stop — which is a much harder problem than personal discipline. ## ✅ Action Items #### DevRel - Developer advocacy: consider referencing this survey data in talks/content about sustainable AI-assisted development practices. - Education: could inform a workshop on healthy AI-coding workflows (time-boxing, deliberate review practices) drawing on Dr. Tolinski's concrete suggestions. - Developer experience: N/A directly — not about tooling UX. - Developer marketing: N/A. - Community management: worth monitoring for community sentiment around AI-coding burnout in WordPress/dev communities Courtney tracks. - Internal/external/cross-project: N/A. #### OSPO - Policy & Governance: N/A. - License Compliance: N/A. - Security Management: N/A. - Software Inventory: N/A. - Developer Support: relevant — burnout/mental-health support resources for contributor communities could reference this data. - Community Engagement: N/A. - Strategic Alignment: N/A. #### Community - Maintainers: directly relevant — several interviewees are maintainers (Adam Elmore, Mark Erikson) describing burnout and coping strategies specific to maintainer workloads. - Contributors: relevant to onboarding conversations about sustainable AI-assisted contribution pace. - Sponsors: N/A. - Ecosystem: worth flagging in cross-project mental-health/sustainability conversations. #### WordPress - Core: N/A. - Core Performance: N/A. - Core AI: directly relevant — as WordPress Core AI initiatives roll out AI-assisted contribution tooling, this data is a useful input on workflow/burnout design. - Core Program Team: N/A. - Mobile: N/A. - Playground: N/A. - Community: N/A. - Training: potential source for a "healthy AI-assisted development" training module. - Marketing: N/A. - Support: N/A. - Design: N/A. - Documentation: N/A. - Photos: N/A. - TV: N/A. - Polyglots: N/A. - Accessibility: N/A. - Meta: N/A. - CLI: N/A. - Hosting: N/A. - Tide: N/A. - Openverse: N/A. - Themes: N/A. - Plugins: N/A. #### The WPOCC - Programs: N/A. - Projects: N/A. - Funding experiments: N/A. ### FAIR - Technical independence: N/A. - Policies and covenants: N/A. - Productized governance artifacts: N/A. ## 🧪 Hypotheses - H1: Developers who regularly prompt past their intended stopping point will show measurably worse sleep-quality metrics (via wearable/actigraphy data) than matched developers who reliably stop on time. - H2: The correlation between AI-agent count and both higher enjoyment and higher sleep disturbance reflects two distinct populations (deeply engaged/energized vs. compulsively overextended) rather than a single dose-response relationship — testable by clustering respondents on additional variables (job security, autonomy, workload). - H3: Workplace autonomy (ability to set one's own schedule/boundaries, as in Aaron Francis's case) moderates the relationship between AI usage intensity and burnout risk more strongly than usage intensity alone. - H4: Self-reported skill diminishment among heavy AI-agent users is confounded by supervisory-role blindness — i.e., some heavy users may be unaware of skill loss precisely because they've shifted to review-only work (the video raises this as an open question rather than resolving it). ### 📊 Methodology - H1: within-subject or matched-cohort study pairing self-reported prompting-stop behavior with objective sleep tracking data. - H2: cluster analysis of the existing survey data using job security, autonomy, and workload variables alongside agent count. - H3: multi-site survey stratified by employment type (self-employed vs. managed/quota-based roles) measuring both usage intensity and burnout outcomes. - H4: skills assessment (objective coding tasks) compared against self-reported skill perception, stratified by agent count/review-vs-write ratio. - **Note on the source survey's own methodology:** self-selected respondent pool (~1,300 people who chose to take a public survey about AI coding's effects on their mental health) — likely skews toward people with strong opinions/experiences on the topic in either direction, a limitation the video states explicitly and repeatedly. ### ❓ Questions - Does the "more agents = less felt skill diminishment" finding reflect genuine skill retention through more active orchestration, or reduced self-awareness of skill loss — and how would you design a study to tell the difference? - What specific organizational policies (not just individual habits) most effectively protect developer wellbeing under AI-accelerated production pressure? - How does the "variable reward" dopamine mechanism described here compare quantitatively to established behavioral-addiction models (e.g., gambling, social media) — is AI prompting meaningfully different in intensity or just a new instance of the same pattern? - Given the survey is ongoing and self-selected, how might its conclusions change (or not) as the sample grows and diversifies? ### 🔮 Future Research - A longitudinal follow-up on the same survey population to see whether burnout/skill-diminishment trends worsen, stabilize, or reverse as developers adapt. - Comparative research on organizational policies (mandatory offline hours, agent-count caps, review-first workflows) and their measured effect on developer burnout in AI-heavy teams. - Independent replication of the "variable reward" addiction framing using established behavioral-psychology instruments rather than self-report alone. ### 💡 Implications - **Practice:** developers and teams adopting heavy AI-agent workflows should proactively build in time limits, offline boundaries, and deliberate human-review steps rather than assuming productivity gains are costless. - **Policy:** employers/managers who expect AI-driven output increases should be aware that the "saved time" often converts to more decisions and supervision, not rest — and that expecting sustained maximum output risks measurable burnout and skill/enjoyment decline in their teams. - **Security:** indirectly relevant — burnout and reduced skill engagement could affect code review quality and security diligence, though this isn't addressed directly in the source. - **Education:** useful teaching material for discussions on sustainable AI-assisted development practices and recognizing early burnout signs. - **Funding:** N/A — no funding/business model angle in this source. ### 🧠 Learning Styles Assessment - **Reading level estimate:** approximately Flesch-Kincaid grade 8–9 for the spoken transcript (conversational register, short clauses, moderate technical vocabulary). #### 👀 Visual learners - Present: video format includes on-screen charts of survey data (implied by "See the charts here" link) and interview footage. - Improve: a companion written article with the actual chart images embedded would help non-video learners. #### 🔊 Aural learners - Present: strong — full video/audio format with multiple voices (interviews) rather than a single narrator. - Improve: N/A, this format already serves aural learners well. #### 🗣️ Verbal learners - Present: strong use of analogy (slot machine/variable reward, musical instrument skill atrophy) to explain concepts. - Improve: a glossary of terms (neuroplasticity, somatization, variable reward system) for viewers unfamiliar with the psychology vocabulary. #### 🧩 Kinesthetic learners - Present: concrete, actionable boundary-setting tips (time limits, lights-off schedule, no prompting apps on phone). - Improve: a downloadable self-assessment checklist based on the survey questions. #### 🧠 Logical learners - Present: strong — data presented with explicit correlation/causation caveats and stated sample limitations throughout. - Improve: an explicit summary table of all stats (46%, 58%, 17%, 65%, 59%, 54%) with their caveats would serve this group further. #### 🧑‍🤝‍🧑 Social learners - Present: strong — multiple named interviewees, explicit call for viewer comments/experiences at the end. - Improve: N/A, already well-served. #### 🧘‍♀️ Solitary learners - Present: the ongoing survey invites self-reflection and individual data contribution. - Improve: N/A. #### 🌿 Naturalistic learners - Present: Not applicable to this source — no environmental/sustainability content. - Improve: N/A. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: not directly assessable — captured content is a scraped transcript, not the original video page markup. - Improve: N/A from available data. #### Keyboard navigation - Present: not assessable from transcript content. - Improve: N/A. #### Contrast and readability - Present: not assessable — no visual/CSS data captured. - Improve: N/A. #### Text resizing and reflow - Present: not assessable from captured content. - Improve: N/A. #### Media and captions - Present: the video has machine-generated captions/transcript (the source of this analysis), which is itself a positive accessibility signal. - Improve: human-reviewed captions (auto-transcripts often mis-transcribe technical terms, as seen with "Anthropic's powerful Fable model" — likely "Claude" mis-heard) would improve accuracy for deaf/hard-of-hearing viewers. #### Forms and interactions - Present: the linked survey (Google Form) is a simple, generally accessible format. - Improve: N/A, not independently assessed here. #### Standards reference - Insufficient captured data to map against WCAG 2.2 AA criteria; the auto-transcript's apparent mis-transcriptions (see Media and captions) would be the first thing to check against 1.2.2 (Captions). ### 🌍 DEIB Assessment #### Cultural breadth - Present: interviewees span different life situations (parent with young kids, established open-source maintainer, self-employed founder) though all appear to be within the same North American tech-industry milieu. - Improve: broader geographic/cultural representation in future survey waves, given AI-coding pressure likely varies by labor market and employment protections outside the US. #### Inclusive language - Present: plain, accessible language; mental health topics discussed with clinical accuracy (via the licensed psychologist) rather than stigmatizing language. #### Neurodiversity support - Present: chaptered structure (per the video's own timestamps) aids navigation; conversational pacing may be dense for some viewers. - Improve: a written summary/transcript with the same chapter structure would help viewers who process text better than audio. #### Representation - Present: features a named psychologist explaining mechanisms rather than only lay speculation; includes a range of career stages and family situations among interviewees. - Improve: N/A. #### Psychological safety - Present: notably strong — the video includes an explicit crisis-resources disclaimer (ADAA.org, 988 in the US) before discussing burnout and mental health content, and frames struggling developers with empathy rather than judgment. ### ✅ Logical Validity #### Strong reasoning - **Correlation-causation discipline** (sleep/stopping-point data): explicitly states "it doesn't mean that AI prompting inherently wrecked their sleep... they did show up together in the data" rather than overclaiming causation from a correlation. Location: sleep-data segment. This is careful, honest handling of observational survey data. - **Explicit acknowledgment of thin data** (agent-count vs. skill diminishment): rather than drawing a conclusion from a counterintuitive finding, the video states plainly that the 4-5-agent group is too small to draw conclusions from. Location: agent-count segment. This is a model of restraint that many data-driven video essays skip. - **Careful causal hedging on employment data**: explicitly notes unemployment rose across all job categories in the cited window, not just AI-exposed ones, undercutting a naive "AI caused this" reading before it can be assumed. Location: closing structural argument. Strengthens the argument's credibility by resisting the more dramatic (but less supportable) causal claim. #### Fallacies - **Misleading vividness (cold open)** — the video opens with the single most extreme case in its entire dataset (a developer hospitalized after an all-night prompting binge) before any systematic data is presented. This primes viewers to interpret the more moderate survey findings (e.g., 46% overshoot their stopping point) through the lens of a hospitalization-level outlier. Location: 00:00–01:08. Fix: explicitly frame the anecdote as an extreme outlier relative to the survey's actual central tendencies, rather than letting it stand as an unqualified hook. - **Downstream generalization risk (self-selected sample)** — the video itself caveats its sample honestly, but the specific percentages (46%, 58%, 59%, 54%, 65%) are exactly the kind of clean statistics likely to be quoted elsewhere (social media, press coverage) without the self-selection caveat attached, functioning as a latent hasty-generalization risk even though the source doesn't commit the fallacy itself. Location: throughout the stats segments. Fix: pair any external citation of these numbers with the sampling caveat every time, not just in the original video. #### Tone of voice - Empathetic, measured, and personally invested (the host discloses his own experience with the same pull); balances first-person narrative, expert authority, and data without either fear-mongering or dismissing the real productivity benefits of AI tools. ## 🛡️ DARVO Analysis - Not applicable in the classic sense — this video is not a response to an accusation, and no party denies wrongdoing, attacks an accuser, or reverses victim/offender roles. Included for completeness: the video does explicitly reject a *potential* victim-blaming frame — it argues that "healthy prompting can't be carried by personal discipline alone" and that workplace structures, not just individual willpower, determine whether boundaries are achievable. This is worth noting because it preempts a common move in productivity/wellness discourse where systemic pressures get reframed as individual failures to self-regulate; the video explicitly declines to make that move. ## 🎯 Conclusion - **Bottom line:** original survey data from nearly 1,300 developers shows real, measurable associations between heavy AI-assisted coding and worse sleep, higher perceived pressure, and diminished skill/enjoyment for a majority of respondents who report those effects — driven by a psychologically plausible variable-reward mechanism — but the video's most important claim is structural: individual coping strategies only work within workplace cultures that permit stopping, so this isn't purely a personal-discipline problem. - **Practical next steps for readers:** consider taking the still-open survey; if using AI agents heavily, adopt concrete boundaries (time limits, device-level restrictions, offline rhythms) similar to those described; and if in a position to influence workplace policy, treat "AI lets us go faster" as a question ("should we?") rather than an unexamined mandate. ## 🔄 Refactor Appendix — The True Cost of AI Coding > 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 prompting works like a slot machine on the brain > [!quote] ID:rw_aicoding_variablereward > "it's a variable reward system, which is more like a slot machine... that in turn induces dopamine, which feeds this addictive cycle." **Context & use**: - Dr. Courtney Tolinski, ~06:42. - The core explanatory mechanism for the whole video's findings; strong, reusable framing for why AI-coding sessions are hard to stop. **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://)  #### [H2] Overshooting stop time strongly predicts worse sleep > [!quote] ID:rw_aicoding_sleepstat > "Of the people who kept going past their stopping point, 58% said their sleep had changed... Of the people who actually stopped when they meant to, just 17." **Context & use**: - Survey data, ~06:03–06:29. - Strongest single relationship in the survey; reusable data point, with explicit correlation caveat attached. **Tags**: #xref #addendum **Highlight**: [view highlight](readwise://)  #### [H3] Individual discipline can't fix structural overwork pressure > [!quote] ID:rw_aicoding_structural > "healthy prompting it can't be carried by personal discipline alone, right? The worker, the workplace, all of these things matter" **Context & use**: - Closing argument, ~22:20. - The video's real thesis; reusable framing against individual-blame narratives in burnout discourse generally, not just AI coding. **Tags**: #xref #articleseed **Highlight**: [view highlight](readwise://)  ### 🧠 Permanent Candidates (H4) #### [P1] Boundary-setting advice needs a workplace-autonomy caveat **Claim** (one sentence, Courtney's voice): Personal-boundary advice for AI-coding burnout (time limits, offline hours) only transfers to people who have Aaron Francis's level of workplace autonomy, so it should always be paired with a structural caveat rather than presented as universally actionable. **Evidence pointer**: [[#H3]] **Scope & conditions**: Applies to burnout/productivity advice generally, not just AI coding; breaks down as a critique for advice explicitly scoped to self-employed/high-autonomy audiences. **Links** (≥1 with annotation): - [[Burnout and workplace structure]] — because extends - [[AI-assisted development]] — because applies-in developer wellbeing discussions  ### ✏️ Fleeting Queue - [F1] Compare this survey's methodology/caveats against the Stanford AI Index employment data it cites directly — worth reading the primary source. #xref - [F2] Watch for follow-up coverage as the Syntax.fm survey stays open and the sample grows — check if findings hold. #xref ### → Things 3 (max 2 actions) - [ ] Consider taking the open developer mental-health survey (https://forms.gle/TvtjgtU7UPnK7DC9A) — context: personal AI-coding workflow reflection #things  📄 **Full text:** [[Full Document Contents/Articles/The True Cost of AI Coding.md|The True Cost of AI Coding by youtube.com]] ## Highlights - past the point in which they [music] were meant to stop. Not once in a while, often, or even every single day. And I personally have felt this same pull. So, I took that group of people and looked at their sleep. Of the people who kept going past their stopping point, 58% said their sleep had changed since they started using AI heavily, moderately, or more. Of the people who actually stopped when they meant to, just 17. And that's the strongest relationship in ([View Highlight](https://read.readwise.io/read/01m005szg546wggaxyd73q6mn9)) - Note: Strongest relationship in the survey: overshooting intended stop time correlates with much worse sleep (58% vs 17%), explicitly caveated as correlational. - Tags: fallacy/downstream-generalization-risk, atom/literature/h1 - this entire survey, [music] by the way. And And it doesn't mean that like AI prompting inherently wrecked their sleep. It could easily, you know, run the other way or be something else entirely. But, they did show up together in the data. >> As you can imagine, it's a variable reward system, which is more like a slot machine, right? Where sometimes you're getting a reward, but most of the time you're not. Right? And And there's this like anticipation build-up that keeps you motivated to continue to press that ([View Highlight](https://read.readwise.io/read/01m005ssgy3gbc3cz8az31g8b3)) - Note: Dr. Courtney Tolinski's variable-reward/slot-machine explanation for why AI prompting is addictive -- core explanatory mechanism for the video's findings. - Tags: atom/zettel/z1, atom/literature/h1 - do we even need to read the code if it works, right? But what does it mean to feel like a programmer? Like what happens to enjoyment and skill and that identity that [music] you can have for being a developer? We asked developers, do you feel like your actual coding skills are sharpening, holding steady, [music] or diminishing? 59% put their own skills on the diminishing side of things. As well as how much genuine enjoyment or flow do you get from coding now compared to ([View Highlight](https://read.readwise.io/read/01m005t2yr4sxncnedvycz65qp)) - Note: 59% of surveyed developers feel their coding skills are diminishing from heavy AI use. - Tags: atom/literature/h2 - that unemployment went up across every kind of job in that window. And it actually went up more for the jobs least exposed to AI instead of [music] the most. So, this isn't proof that AI took those jobs. But if you're like 23, and you feel like the ladder is possibly being pulled up behind you, you're going to be pushing hard, right? You're probably going to be pushing past your [music] set boundaries and limits that other people who have those abilities might be in more control of. So, healthy prompting it can't be carried by personal discipline alone, right? The ([View Highlight](https://read.readwise.io/read/01m005t9sdmvnnf89ghyv82kqw)) - Note: The video's real thesis, carefully preceded by a correctly-hedged causal claim about employment data: structural/workplace factors, not just individual discipline, determine whether healthy AI-coding boundaries are achievable. - Tags: atom/literature/h1, atom/zettel/z1