Skill decay is the loss of a trained capability during a period of non-use. It is measured against a retention interval — time since the skill was last practised — not against the calendar. Meta-analytic evidence puts the rate at roughly 0.06 to 0.08 standard deviations per month of non-use, varying widely by task, by person, and by how closely the test resembles the job.
That is the measured quantity. It is not what "the half-life of skills is five years" means, and the gap between the two is where a lot of L&D and workforce policy quietly goes wrong.
The number everyone quotes, and the number the research produced
"The half-life of a skill is five years" — or two and a half, depending on whose slide you are looking at — appears in capability strategies, credential expiry policies and board papers. Follow the citation chain and it usually terminates at another article rather than at a dataset. It reads like a finding. It behaves like a metaphor.
The actual literature is large and unglamorous. The foundational meta-analysis, Arthur, Bennett, Stanush and McNelly (1998) in *Human Performance*, pooled 189 independent data points from 52 articles and reported decay as a function of the retention interval: essentially nothing below one day (d = −0.01), a sharp drop within the first week (d = −1.01 at one to seven days), and d = −1.27 beyond a year of non-use.
The 2025 update, Tatel and Ackerman's procedural skill retention meta-analysis in *Psychological Bulletin*, screened 457 sources and coded 1,352 effect sizes. Its headline is a slope, not a half-life: standardised performance differences grow by about 0.08 per month on accuracy measures and 0.06 per month on speed and mixed measures, and the gains from initial training are typically given back between one and two and a half years when the skill is rarely used. The open dissertation version carries the full moderator analysis.
Note what the x-axis is. It is time since the skill was last performed. An underwriter who prices risk every working day has a retention interval of one day, whatever year she qualified. A half-life framing measures from the wrong origin.
Myth 1: Skills expire on a schedule
The claim: capability has a shelf life, so competence should be re-certified on a fixed cycle.
The evidence: decay tracks use, not elapsed time. Both meta-analyses above model performance against non-practice intervals, and the steepest losses in Arthur et al. occur in the first days of non-use rather than accumulating smoothly across years. A cycle set by the calendar re-tests the people who never stopped practising — the cheapest possible false negative for your budget — while leaving genuinely dormant skills untouched until the next anniversary.
Myth 2: There is a single rate you can plan around
The claim: pick the number, apply it organisation-wide.
The evidence: variance is the finding, not noise around it. In Arthur et al., cognitive tasks decayed markedly faster than physical ones (d = −1.15 versus −0.75), and the single largest moderator was not task type at all but similarity of retrieval conditions — a difference of 1.13 standard units, nearly three times the physical-versus-cognitive gap. How closely the assessment resembles the real task moves the measured decay more than what kind of skill it is.
Narrow the field to one high-stakes domain and the spread persists. A systematic review of advanced life support retention (Yang et al., Resuscitation, 2012) covering 11 studies concluded that ALS knowledge and skills decay by six months to a year after training — but observed deterioration anywhere from six weeks to two years, with skills consistently decaying faster than knowledge. If a single clinical procedure with standardised training cannot produce one number, a whole economy certainly cannot.
Myth 3: Market churn and personal decay are the same problem
The claim: skills are changing fast, therefore your people's skills are expiring fast.
The evidence: these are two different measurements with two different fixes. The churn side is real and well documented. Shifting Skills, Moving Targets — the Burning Glass Institute, Emsi Burning Glass (now Lightcast) and BCG analysis of more than 15 million US job postings from 2016 to 2021 — found that over a third of the top 20 skills requested for the average job had changed, with roughly one in five an entirely new requirement, and nearly three-quarters of jobs changing more between 2019 and 2021 than in the preceding three years.
It also found the change is wildly uneven: in the fastest-moving roles close to 80% of the top 20 skills were new or substantially re-weighted, while warehouse and janitorial roles saw 15% or fewer change. The World Economic Forum's Future of Jobs Report 2025 puts employer expectations at 39% of core skills changing by 2030, down from the 44% projected in 2023 — and is explicit that this is a survey of employer expectations. Its own workforce model is more useful than the headline: of 100 representative workers, 41 need no significant training at all, 29 are upskilled in role, 19 reskilled and redeployed, and 11 need training they cannot access.
Your engineer's Python did not rot. The role started asking for something else. Non-use decay is fixed by practice; demand churn is fixed by re-specifying the job and re-testing against the new specification. Conflating them produces training spend aimed at the wrong half of the problem.
What to do instead: measure the use rate
A defensible re-assessment policy replaces one organisation-wide interval with a per-skill rule.
- Set the interval per skill, not per person. Ask how often this specific capability is actually exercised in the role. Daily-use skills need monitoring, not re-testing; quarterly-or-rarer skills are where decay accumulates.
- Separate the two questions explicitly. "Has this person's capability decayed?" and "has this role's requirement changed?" need different evidence and different budgets.
- Prioritise by consequence, not recency. Low-frequency, high-consequence skills — incident escalation, fraud triage, safety overrides — earn the shortest intervals regardless of when training happened.
- Test under job-like conditions. Given that retrieval similarity was the largest moderator in the decay literature, scenario-based and work-sample formats give a truer decay reading than recognition-format quizzes. Scoring judgement-based responses at scale is the practical obstacle, which is where AI-graded scenario assessment with integrity bands, as in AssessAll's scenario formats, makes frequent low-stakes checks affordable.
- Report knowledge and performance separately. They decay at different rates; a combined score hides the one that moved.
- Re-anchor the standard when the job changes. The SIOP Principles for the Validation and Use of Personnel Selection Procedures treat substantive change in the work as the trigger for revisiting validation evidence — job change, not the passage of time, is what invalidates a standard.
- Instrument before you legislate. Run a short baseline, re-measure a sample at three, six and twelve months, and derive your cadence from your own curve. Pay-as-you-go credits at ₹30 / US$0.50 per assessment make a sampling design like this cost a fraction of a re-certification programme built on a borrowed number.
When a fixed expiry date is the right answer
Sometimes it is, and pretending otherwise is not rigour. A regulator that mandates a two-year recertification sets the interval, and no internal evidence overrides it. Bright-line expiry is also the sensible default where the population is too large to instrument, where audit needs a rule anyone can verify, or where a skill is exercised so rarely that any interval is effectively "always dormant". The honest version is to say that the date is an administrative convention chosen for defensibility — not to dress it up as a half-life.
The takeaway
Skills do not have a half-life; they have a use rate, and the measured evidence describes decay per month of non-use with enormous variance by task, condition and person. Build your re-assessment cadence on how often each capability is actually used and what it costs when it fails — and if you are going to publish a number, publish your own.