A critical review of what the evidence actually supports — and of the popular claims that do not survive reading the primary sources.
Compiled 26 July 2026 · 119 primary sources · every headline figure traced to source · contested and unverifiable claims flagged throughout
The short version
Nine things the evidence supports
Ask why chronic stress leaves people unable to function and you will usually be told a story about cortisol exhausting the adrenal glands and willpower running down like a battery. Both halves are wrong. What replaces them is more interesting.
+60%
Cellular energy expenditure under chronic glucocorticoid exposure — chronic stress is hypermetabolic, not depleted
Bobba-Alves 2023 · human fibroblasts, in vitro
d = 0.06
Ego depletion across 36 preregistered sites. Willpower is not a depletable resource
Vohs 2021 · N = 2,463 · BF ≈ 4:1 for the null
16%
Report full recovery seven years after clinical exhaustion disorder — while 87% are back at work
Glise 2020 · N = 217
2 of 80
Prespecified outcomes significant in a 32,974-employee workplace wellness RCT — both self-reported
Song & Baicker 2019 · 160 worksites
0
Empirical validations of the popular "12 stages of burnout" in the entire indexed literature
Europe PMC + PubMed, positive controls used
18.5%
Of fatigue presentations in primary care are depression. Serious somatic disease is no more common than in untired patients
Stadje 2016 · 26 studies
And one competing explanation that has to be excluded first. Chronic restriction to six hours' sleep produces deficits equivalent to two nights of total sleep deprivation — and subjective sleepiness saturates while objective impairment keeps accumulating. People restricted to six hours cannot feel the difference between that and four. Any causal model of stress-related fatigue should control for sleep duration before reaching for the HPA axis. Van Dongen 2003, Sleep 26:117–126
Folk model vs evidence
Six claims that do not survive the primary sources
These are not fringe beliefs. Each is repeated in mainstream health writing, corporate wellbeing material, or both.
"Chronic stress exhausts your adrenal glands"
A systematic review screening 3,470 articles and including 58 studies found "an almost systematic finding of conflicting results" and concluded verbatim that "adrenal fatigue is still a myth." Real HPA dysregulation exists and real adrenal insufficiency exists. The adrenal cortex does not run out.
Cadegiani & Kater 2016, BMC Endocr Disord 16:48
"Willpower is a battery that runs down"
Ego depletion failed two enormous preregistered multi-lab tests: d = 0.04 across 23 labs and d = 0.06 across 36 sites. The better account is that stress raises the price of effort rather than removing the capacity — a distinction that changes which interventions make sense.
Hagger 2016; Vohs 2021, Psychol Sci 32:1566
"Burnout progresses through twelve stages"
Zero hits for the phrase in Europe PMC or PubMed against validated positive controls. No instrument, no staging criteria, no prospective test. Longitudinal cohorts recover trajectory classes distinguished by severity and treatment response, never stage position — and find exhaustion and vigour behave as independent constructs.
Dunford 2012; Mäkikangas 2012; Hätinen 2009
"High cortisol destroys your deep sleep"
Acute cortisol administration in healthy volunteers increases slow-wave sleep and suppresses REM. But this does not generalise: in Cushing's syndrome delta sleep is reduced (5.8% vs 14.0%), and ACTH-independent Cushing's still disrupts sleep. The claim is wrong about acute pharmacology and roughly right about chronic states.
Friess 1994; Friedman 1994; Shipley 1992
"Resilience training reduces workplace burnout"
A cluster-randomised trial of 32,974 employees found 2 of 80 prespecified outcomes significant, both self-reported; all clinical, spending and employment outcomes null. A second RCT found participants were already healthier and cheaper before enrolling — apparent wellness ROI can be entirely a sorting artefact.
Song & Baicker 2019/2021; Jones 2019
"Ashwagandha lowers stress"
Pooled effects look spectacular (SMD −1.6) but the literature contains no negative trials at all (Egger's p < 0.001), mean trial size ~51, GRADE low-to-very-low. Most tellingly, the one review reporting both found cortisol fell while perceived stress was flatly null (p = 0.40). LiverTox rates it a likely cause of clinically apparent liver injury.
Marchi 2025; Albalawi 2025; LiverTox NBK548536
Synthesis
A defensible causal model
Boxes with a dark left edge are well supported; pale edges are weak or contested. Note what is missing: cortisol appears nowhere as a load-bearing pathway, because its measured associations sit at r ≈ 0.10–0.15 and two decades of searching found no burnout biomarker.
Sleep lossInitiation and continuity6 h × 14 nights ≈ 2 nights total deprivation
↓
Metabolic cost of allostasisThe bill is paid out of growth, maintenance and repairstrong at the cellular level (+60%) · unproven whole-body — the field's central open question
↓
Low-grade inflammationSickness behaviour; dopamine synthesis impairedweak at population level, z ≈ 0.10
Raised effort priceEffort as a computed cost, not a fuel gaugelPFC glutamate accumulation → low-cost bias
Impaired prefrontal functionReflective → reflexive controlacute small · chronic broad (g −0.36 to −0.53)
↓
Motivational withdrawal"I can't face it" — a raised price, not an empty tank
↺ Feedback: the strain → stressor path is larger than the stressor → strain path (Guthier 2020, k = 48, N = 26,319). Once running, the state maintains its own conditions.
What degrades
Acute stress is small and fractionated. Chronic exhaustion is broad.
The single most useful clinical distinction in this literature. Acute stress nudges a few functions; clinical burnout degrades the fluid ones while leaving crystallised ability entirely intact — which is why exhausted people can hold a conversation competently and then fail to plan their week.
Cognitive effect sizes: acute stress vs clinical burnout
Hedges' g with 95% CI · negative = impairment · tap or hover any point for detail
↔ scroll the chart sideways, or use the table view
Read the intact rows. Inhibition overall is null under acute stress (g = −0.076) — the frequently repeated "acute stress enhances response inhibition" is a post-hoc moderator result on roughly five degrees of freedom. Memory consolidation is the only positive effect: stress strengthens what has just been laid down while degrading access to what is already stored. That retrieval deficit, worse for emotional material, is the mechanism of blanking under pressure — and of recalling everything perfectly twenty minutes afterwards.
Cortisol is a poor proxy for cognitive impact. In the same meta-analysis, cortisol reactivity did not moderate the stress effect on executive function, and exogenous cortisol administration produced no comparable impairment (g = 0.030, p = .495, versus stress g = −0.151, p = .005). The hormone everyone measures is not the variable that predicts the outcome anyone cares about.
The chronic tail
Depression resolves in months. Exhaustion does not.
The most under-communicated finding in this review, and the strongest argument for acting well before the clinical threshold.
Symptom persistence after clinical exhaustion disorder
% above clinical threshold · Swedish cohorts, N = 232 (18 mo) and N = 217 (7 y)
↔ scroll the chart sideways, or use the table view
Anxiety and depression are plotted as unconnected endpoints because intermediate values were not reported — no interpolation is implied. The burnout curve flattens at around 18 months and stays flat: at seven years, 31% were still clinically judged to have stress-related exhaustion, 46% still reported extreme fatigue and 73% reduced stress tolerance — while 87% were not on sick leave. At ten years, symptoms of burnout, anxiety and depression "remained stable from the 1- to the 10-year follow-up," and 73% had changed workplace against a 25.5% population base rate.
Functional recovery arrives long before symptomatic recovery. That is the honest reading of these cohorts, and it cuts both ways: reaching the clinical threshold has consequences that outlast the sick note by years, but reduced stress tolerance is plainly compatible with a working life. The 73% who changed workplace look less like casualties than like people who made an accurate adjustment.
Interventions
What actually helps, ranked honestly
Effect sizes flipped so that positive always means benefit. Colour marks the strength of the evidence base, not the size of the effect — a large effect from weak trials is still weak.
Intervention effect sizes with 95% confidence intervals
Standardised mean difference · tap or hover any point for detail · intervals crossing zero are non-significant
Strong Moderate Limited Null / insufficient
↔ scroll the chart sideways, or use the table view
Three things to read off this chart. First, the largest point estimate (CBT for occupational stress, d = 1.16) has the widest interval and rests on seven studies — size and certainty are not the same axis. Second, everything from micro-breaks downward has an interval that touches or crosses zero. Third, the two interventions with both a strong evidence base and a solid effect — CBT-I and behavioural activation — are structured behavioural protocols, not wellbeing products.
The ordering that follows
Do first
Why
Fix the sleep disorder CBT-I; investigate apnoea if plausible
The only strong guideline recommendation in this review (AASM). g ≈ 0.98 for insomnia severity — but note g = 0.16 for total sleep time: CBT-I fixes quality, not hours.
Move, 3–5× a week Resistance or aerobic, moderate-to-vigorous
SMD −0.42 to −0.62 against active controls. Steepest returns on the first increment from nothing; benefit appears below public-health guideline volumes.
Screen for and treat depression
Highest base rate of any treatable cause of fatigue (18.5%), and the strongest treatment evidence in the review.
A structured behavioural protocol Behavioural activation
SMD 0.67, and it works self-guided (0.36). The highest evidence-to-complexity ratio available — and it targets the effort–reward structure rather than trying to add capacity.
Protect detachment structurally
Best-supported recovery construct (r ≈ −0.42 with fatigue). Must be structural, not aspirational — see the recovery paradox below.
The recovery paradox is the design constraint on all of this. Job demands predict lower detachment (r̄ = −0.25 across 60 studies, N = 28,507). The people who most need recovery are least able to initiate it — because activation keeps work cognitions accessible, because depletion makes effortful recovery harder to start, and because connectivity removes the boundary. This predicts that telling depleted people to "recover better" will fail, and it is the strongest argument in the literature for making recovery a structural feature rather than a personal discipline. Sonnentag 2018
Recovery that removes exhaustion is not recovery that restores energy. Detachment predicts reduced fatigue (ρ = −0.39) but barely predicts vigour (ρ = +0.14). It is control (0.31) and mastery (0.29) that generate vigour. Collapsing on the sofa and learning an instrument are not interchangeable, and the literature can tell them apart. Bennett, Bakker & Field 2018, N = 26,592
Negative findings
Overhyped or unsupported
Named explicitly, because the alternative is that they keep circulating.
Sleep hygiene as a standalone intervention — the AASM recommends against it. The active ingredients of CBT-I are stimulus control and sleep restriction, which are demanding and initially worsen daytime sleepiness. The easy component is the one that gets promoted.
Mindfulness as a general-purpose intervention — moderate for anxiety, depression and pain; low or insufficient for stress, sleep, attention and mood; d = −0.004 versus an evidence-based treatment. Under 25% of trials assess adverse events, and pooled adverse-event prevalence reaches 33% in observational studies. (Credit where due: MBSR was noninferior to escitalopram for anxiety disorders in a 276-person randomised trial.)
Rhodiola — the confirmatory double-blind RCT found vitality favoured placebo (MD −17.3, p = 0.011). It made fatigue worse.
Magnesium for stress — the flagship review states that no study administered a validated measure of subjective stress, none recruited depleted participants, and every placebo-controlled study showed significant placebo effects.
Omega-3, vitamin D and B vitamins in non-deficient people — null in megatrials (vitamin D: HR 0.97 across N = 18,353 over 5.3 years). The VITAL omega-3 arm found a significant increase in depression risk.
"Caffeine 6 hours before bed costs an hour of sleep" — traceable to an n = 12, manufacturer-funded study with no polysomnography, whose diary result was non-significant (p = 0.08) and which explicitly found no timing gradient. The real figure: caffeine costs ~45 min of total sleep, recovered at ~2.8 min per extra hour before bed.
"Alcohol causes REM rebound" — the meta-analysis finds whole-night REM reduced by 2.8%, not rebounded, with total sleep time unaffected. Which is exactly why people don't notice.
"The four-day week cut burnout by 71%" — the 71% is the proportion of individuals whose score moved down at all, in an uncontrolled, self-selected, self-report pilot with no randomisation.
"Micro-breaks boost productivity" — the pooled performance effect is non-significant (d = 0.16, CI crosses zero) with evidence of small-study bias, and longer breaks did better. Take breaks for recovery, not output.
Wellness apps — g = 0.09 (ns) against a placebo app; real-world median 30-day retention 3.3%. For breathing-exercise apps specifically: 0.0%.
"Loneliness kills as much as 15 cigarettes a day" — not a figure that appears in Holt-Lunstad's papers.
Reference
Every key number
Filter by keyword or category. Figures marked null are non-significant and are included deliberately — the negative results are half the story.
Finding
Value
Source
Method
How to read any number in this field
Three structural problems
Control-group inflation. Effect sizes collapse as controls get more active. Mindfulness: d = 0.55 vs no treatment → 0.35 vs non-specific active control → 0.23 vs specific active control → −0.004 vs an evidence-based treatment. Whenever you see an impressive effect size, the first question is what it was compared against.
Self-report and unblinding. Almost every outcome here is a questionnaire completed by an unblinded participant. The honest counterweight: MetaBLIND found no average difference between blinded and unblinded trial estimates. A caution, not a proof of inflation.
Publication bias, sometimes severe. Trim-and-fill cut the app-for-depression effect by ~60%. The job-strain consortium documented its own: published estimates 1.43, unpublished 1.16.
Claims this review corrected or could not verify
Listed so nobody launders them onward. Roughly one headline claim in three needed narrowing once the primary source was read.
"Burnout does not present the unity expected of a distinct syndrome" attributed to Bianchi 2021 — misattributed. Zero full-text hits. The phrase "no syndromal unity" is from Verkuilen 2021, Assessment. The actual 2021 conclusion is the weaker "burnout problematically overlaps with depression." The CI [.75, .84] is also unverified; the point estimate r = .80 is confirmed.
"Acute stress enhances response inhibition" — overstated. Overall inhibition effect null; the +0.296 is a post-hoc moderator on ~5 df.
"Cortisol increases deep sleep, so the popular claim is backwards" — bounded. True for acute exogenous cortisol in healthy volunteers only. Chronic hypercortisolism reduces delta sleep, and ACTH-independent Cushing's still disrupts sleep, which defeats the "it's CRH drive, not cortisol" attribution.
Fleming 2024 — design confirmed cross-sectional, not randomised. The widely repeated per-intervention details ("volunteering the only positive signal," "mindfulness negative") are not in the abstract and the full text was inaccessible. Cite the headline null; not the details.
Stadje 2016 "identical" somatic disease prevalence — softened. Only 6 of 26 studies had controls; the claim rests on two single studies, one powered at 0.09. The abstract gives 4.3%, the discussion 3.1%.
ICD-11 code "QD85" — not stated on the WHO news release itself. Note also that burnout appeared in ICD-10 in the same category; the 2019 announcement was neither new nor a recognition as a condition.
WHO's "US$1 trillion in lost productivity" — the fact sheet provides no source or methodology. "€617bn cost of work-related stress" is a mis-citation: it estimates work-related depression, and the source report states verbatim that "stress was not used."
West 2016 subgroup point estimates — the paper publishes only an interaction p-value. Any source quoting subgroup effect sizes is fabricating them.
Also unverified: Lehrer 2020 HRV pooled effects; Fan 2025 four-day-week coefficients; van Straten 2018 and Bennett 2018 confidence intervals; Starcke & Brand 2016 pooled estimates; any pooled effect for stress → vigilance decrement; social-support buffering meta-analyses.