The Human Cost. How AI-era work feels, on the record.
One more prompt, you said. When did you actually stop?
Read this first: Atomic Reps is a practice tool for understanding and skill. It is not a health product, and nothing on this page claims it prevents, treats, or reduces burnout, sleep loss, or compulsive use. The page exists because how the work feels is part of the evidence about AI-era engineering, and it deserves the same grading as the rest of this site.
- 01In a Syntax poll of 3,593 developers, 71% said they feel they should be producing more because AI makes it possible, 63% placed their own coding skills on the diminishing side, and 57% said they enjoy coding less than before. Self-selected podcast audience, self-reported: it measures how these developers say it feels, not the developer population.
- 02About 52% of the same respondents said they blow past their intended stopping point often or daily. Of those, 63% reported moderate-or-greater sleep change, against 23% of those who stop on time. The poll's own words: association, not evidence of cause and effect. And the most common sleep answer across everyone is still no change at all, at 35%.
- 03The counterweight, printed on purpose: DORA's 2025 survey of 4,867 technology professionals measured burnout directly and found no relationship with AI adoption. What did rise with AI adoption was software delivery instability.
- 04Burnout itself is defined by the WHO as an occupational phenomenon from unmanaged chronic workplace stress, not a medical condition. No peer-reviewed prevalence estimate for software developers using a validated instrument exists; the circulating percentages are vendor survey items.
- 05The best-evidenced mechanism nearby is behavioral, and it is about social media, not coding: across a million-plus posts, people timed voluntary actions to maximize intermittently arriving social rewards. Whether prompting works the same way is untested, and this page says so rather than assuming it.
The sentiment, printed with its grade on its face.
In mid-2026 the Syntax podcast asked its audience how AI-assisted work was going. It kept collecting. The refresh published on 28 August 2026 covers 3,593 responses gathered between 31 July and 28 August.
This is a self-selected poll of one podcast's audience. Every measure is self-reported. There is no instrument validation and no microdata release. It cannot tell you what developers in general experience.
What it can tell you is what 3,593 developers who wanted to talk about this said. The poll publishes its own methodology note: results are descriptive and exploratory, and its correlations are associations, not causal effects. We hold it to that framing everywhere below.
- 71%say they feel they should be producing more, because AI makes it possible2,535 of 3,593 · often or daily
- 63%place their own coding skills on the diminishing side2,279 of 3,593 · five-point scale · belief about skill, not skill
- 57%report less or somewhat less enjoyment from coding than before heavy AI use2,060 of 3,593
Every figure describes this poll's respondents, not developers in general. The largest survey that measured burnout directly found no relationship with AI adoption. That is section 03.
All three rose against the smaller first sample this page carried until now, by four to six points on roughly triple the responses.
The compulsion questions are why this page exists.
Asked how often they continue prompting past when they meant to stop, 1,867 respondents (52%) said often or daily. Split sleep by that answer and you get the poll's sharpest cohort contrast: 63% of frequent overrunners reported moderate-or-greater sleep change since using AI heavily, against 23% of those who rarely overrun.
The poll's Spearman correlation between overrunning and sleep change is 0.41, its second strongest after overrunning and agent count at 0.42. Its own sentence follows every such number: this is an association, not evidence of cause and effect.
Worse sleep could drive more late prompting. Both could follow from workload. Or from the kind of person who answers a poll like this.
Transcribed from the poll's published cohort split (ai-health.syntax.fm, Aug 28 2026 refresh; mid-scale stopping answers sit in neither group, so the two bars do not cover all 3,593). The poll's own framing travels with the chart: a descriptive cohort difference and an association (Spearman rho 0.41), not evidence of cause and effect. Self-selected sample, self-reported measures.
Two details most retellings drop, kept here because they cut against the tidy story.
The modal sleep answer is still no change: 1,253 of 3,593 (35%), the single largest block on that question even as the tail grew.
And perceived skill decline barely differs between heavy and light overrunners, 64% versus 60%. The skills worry does not track how compulsively people prompt.
A poll this self-selected that still refuses to line up neatly is at least being honestly messy.
Transcribed from the poll's published distribution (ai-health.syntax.fm, Aug 28 2026 refresh, responses collected Jul 31 to Aug 28; N=3,593, self-selected, self-report). The headline tail is the moderate-or-greater share, 1,708 of 3,593 (48%). Printing the tail without the modal no-change answer tells half the distribution, so the whole distribution is what we print.
Among respondents who overrun by their own pull, with no outside pressure, more enjoyment was common (46%). Under outside pressure only, it was rare (12%).
The own-pull cell is 173 people out of the 2,489 the poll can classify, so it is a texture worth naming rather than a finding to lean on. The pressure, though, is the majority experience in this sample either way.
The biggest survey finds no burnout link. We print it first.
If AI-era work were straightforwardly grinding developers down, the largest survey we can cite should show it. It does not. This page prints that on purpose. DORA's 2025 research surveyed 4,867 technology professionals: 90% use AI as part of their work, more than 80% believe it has increased their productivity, and 30% report little to no trust in AI-generated code.
DORA measured burnout as an outcome, defined as feelings of exhaustion and cynicism related to one's work, and compared otherwise-similar respondents at higher and lower AI adoption. The result, in the report's own words: no relationship with burnout, no relationship with friction, and an increase in software delivery instability. Their summary sentence is worth quoting whole: "Despite all the benefits, friction remains unaffected, burnout stays flat, and delivery instability rises unless the surrounding system and culture changes."
DORA's interpretation is the systems one: AI shows up at the keyboard, so keyboard-adjacent outcomes improve, while burnout and friction live in the surrounding organization. That is one reading. Another is that a cross-sectional survey with a validated-sounding but short outcome measure would miss slow-building effects.
Either way, the honest sentence for this page is: the strongest available evidence does not show AI adoption raising burnout, and anyone telling you it plainly does is ahead of the data. The sentiment in the poll above and this null coexist. Feeling squeezed by new expectations is not the same construct as clinical-scale exhaustion.
- BurnoutNo relationship with AI adoption
- FrictionNo relationship with AI adoption
- Delivery instabilityRises with AI adoption
Google DORA, State of AI-assisted Software Development 2025 (2025)
An occupational phenomenon, and a number that does not exist.
The word burnout gets used more casually than its definition allows. The WHO's ICD-11 defines burnout as a syndrome resulting from chronic workplace stress that has not been successfully managed, with three dimensions: energy depletion or exhaustion, increased mental distance or cynicism about one's job, and reduced professional efficacy. It is classified as an occupational phenomenon, explicitly not a medical condition, and the WHO says it should not be applied outside of the occupational context. These two parts are what bring us here. Burnout is about work, and it is not a diagnosis anyone hands out from a survey.
Serious burnout research runs on validated instruments, most commonly the Maslach Burnout Inventory, whose subscales are recognizably the three ICD-11 dimensions. That matters because almost every burnout number circulating in developer media is not that. The widely quoted line that 73% of developers have experienced burnout is a single item from a 2023 JetBrains ecosystem survey of 26,348 developers, and the 60%-of-maintainers-considered-leaving figure is a Tidelift survey of about 400 people.
Real signals of sentiment; not measurements of the syndrome. We searched for a validated-instrument prevalence estimate for software developers and did not find one. That number does not currently exist, so nobody gets to quote it, including us.
What the peer-reviewed software literature does offer: a systematic mapping of 92 studies going back to the early 1990s, showing the field moving from qualitative accounts toward quantitative and machine-learning detection. And one large modeling study: a survey of 3,281 developers at one multinational found organizational culture, climate for learning, sense of belonging and inclusiveness associated with work satisfaction, which in turn was associated with reduced burnout.
Associations inside one company, and the authors say so. The consistent shape across this literature is that burnout tracks the surrounding organization, which is the same place DORA's null points.
World Health Organization, ICD-11, QD85 (occupational phenomenon) (2019)Tulili, T. R., Capiluppi, A. & Rastogi, A., Information and Software Technology 155 (2023)Trinkenreich, B., Stol, K.-J., Steinmacher, I. et al., ICSE-SEIP 2023 (2023)
Sometimes the next response is perfect. That is the hook.
Why is one more prompt so hard to put down? The honest answer is that nobody has studied prompting. What exists is an old, solid behavioral literature and one modern application of it, and the bridge between them and coding agents is an analogy we will label as one.
Reinforcement that arrives on a variable schedule produces persistent, high-rate behavior. That is a foundational finding of the animal laboratory, catalogued at book length in 1957. It is about schedules of consequences, not about anyone's neurons, and we state it strictly as behavior. A model's next response is sometimes exactly what you needed and usually not, which is a variable schedule if anything ever was. That sentence is the analogy, and it is untested for developer tooling.
A Nature Communications study analyzed over a million posts from more than four thousand people across social platforms and found posting behavior conformed, qualitatively and quantitatively, to reward-learning principles: people spaced their voluntary actions to maximize the average rate of intermittently arriving social rewards, and an experiment (n=176) reproduced it under manipulation.
The authors' own limitation travels with it: their data do not speak to whether intense use is maladaptive or addictive. So the defensible chain is short: human action timing demonstrably tunes itself to intermittent rewards; AI coding surfaces deliver intermittent rewards; whether the first fact governs the second surface has not been measured.
What we will not do in this section: name-a-neurotransmitter. The behavioral facts stand on their own, and a company selling a practice tool has no business narrating anyone's brain chemistry.
Lindström, B., Bellander, M., Schultner, D. T. et al., Nature Communications 12:1311 (2021)
Effort redistributed is not effort saved.
The pressure number in the poll (71% feel they should produce more because AI makes it possible) implies a mechanism: saved effort returns as raised expectations. Direct evidence for that mechanism is thinner than the feeling is common, and here is what exists.
A survey of 415 software practitioners, analyzed against the SPACE productivity framework, found frequent GenAI users reporting faster task completion and higher output volume, offset by increased code-review burden and persistent cognitive load from verifying AI output, with collaboration patterns unchanged. The authors' reading: perceived productivity gains may be spurious, surface-level acceleration, accompanied by redistributed effort and hidden costs. It is a preprint and entirely self-reported, so grade it as a well-structured version of what practitioners say, not a measurement of what happens.
Beyond that, we found no peer-reviewed study establishing that AI raises the output expected of developers. An ethnography of one company reported in the business press points the same direction, and it is not citable here until the academic paper lands. So the stated gap: the most widely felt claim on this page, that the time AI saves comes back as more expected work, currently rests on self-report. If a real measurement appears, this paragraph changes.
Every source, graded.
The grade column is ours. On this page it matters more than usual: most of what exists on this subject is sentiment, and the grades say which rows are which.
| Source | N | Design | Subject | Grade |
|---|---|---|---|---|
| Syntax poll 2026 | 3,593 | Self-selected online poll | AI-era work, sleep, enjoyment, pressure | Attributed sentiment, not evidence-grade |
| DORA 2025 | 4,867 | Industry survey | AI adoption vs outcomes incl. burnout | Large industry survey, cross-sectional |
| WHO ICD-11 (QD85) | n/a | Classification | Defines burn-out, occupational | Definitional authority |
| Tulili et al. 2023 | 92 articles | Systematic mapping study | Burnout research in software engineering | Peer-reviewed map, no prevalence |
| Trinkenreich et al. 2023 | 3,281 | Survey + SEM, one company | Culture and climate vs burnout | Peer-reviewed; associational, one firm |
| Heath 2025 | 30 papers + 57 items + 7 interviews | Rapid review + consultation | Burnout in open source | Not peer-reviewed; Sentry-funded |
| Ferster & Skinner 1957 | n/a | Laboratory monograph | Schedules of reinforcement | Foundational; animal behavior |
| Lindström et al. 2021 | 1M+ posts; exp n=176 | Computational model + experiment | Reward learning on social media | Peer-reviewed; not about coding |
| Afroz et al. 2025 | 415 | Practitioner survey (SPACE) | Perceived productivity redistribution | Preprint, self-report |
Where this page is weakest.
The centerpiece is a self-selected poll.
People who feel strongly about AI and wellbeing are the people who answer a poll about AI and wellbeing. Every proportion on this page from the Syntax poll describes its respondents, and no reweighting can fix that. We print it because how work feels is worth documenting, not because it measures the population.
Nothing here is causal.
The poll's correlations are associations and say so. DORA is cross-sectional. The burnout modeling study is associational inside one firm. No study on this page can tell you that AI use causes any wellbeing outcome, in either direction.
The strongest evidence points at a null.
DORA's 4,867-person survey found no relationship between AI adoption and burnout. If this page's sentiment data and that null feel contradictory, hold the difference: felt pressure and measured burnout are different constructs, and only one of them has been measured well. The 4,867 is the report's own count, page 103 of PDF v2025.2, where the summary chapter rounds it to nearly 5,000. Cui et al.'s Copilot experiments happen to cover 4,867 developers as well: different studies, identical N, and we never let the two share a sentence without their names.
The prevalence number people want does not exist.
There is no validated-instrument estimate of burnout prevalence among software developers. The circulating figures are vendor survey items, and we name them as such rather than laundering them into statistics.
The mechanism section is an analogy.
Variable-reward behavior is established in animals and demonstrated for social media timing. Nobody has run that study on prompting or coding agents. The parallel is plausible and unmeasured, and the page says which.
Two data sources were excluded on principle.
Two vendor datasets about when developers work (evening commits, weekend activity) would have strengthened the narrative and came from keyboard telemetry of monitored employees. A company that ships coaching-not-surveillance does not cite surveillance, so they are out, and the page is weaker for it on purpose.
The workload mechanism rests on self-report.
That saved effort returns as raised expectations is the page's most intuitive claim and its least measured: one preprint survey and one poll question. We flag it as a stated gap rather than dressing it up.
Nowhere. Atomic Reps is a practice tool for understanding and skill; it does not prevent or treat anything documented above, and we will never claim it does. Two design facts are still worth stating plainly, as facts: the daily rep is one question, and when you answer it, it ends. There is no feed under it. The case for the practice itself, on its own evidence, lives on the hub.
Read the practice evidence on the hubThe receipts.
What developers report
The counterweight
What burnout actually is
- World Health Organization, ICD-11, burn-out (QD85) (2019)
- Tulili, T. R., Capiluppi, A. & Rastogi, A., Information and Software Technology 155 (2023)
- Trinkenreich, B., Stol, K.-J., Steinmacher, I. et al., ICSE-SEIP 2023 (2023)
- Heath, M., Report on burnout in open source (non-peer-reviewed, Sentry-funded) (2025)
The pull, behaviorally
The letter
One study at a time, from issue one.
The Retrieval is this page in instalments: one study worth knowing about, one idea worth a name, and one question you answer from memory in about thirty seconds. Everyone starts at issue one, so nothing in it assumes you read the last one.
An email address. Unsubscribe from any issue.
AtomicReps30″Documented, not diagnosed
The way the work feels is part of the evidence. The practice case stands on its own pages.