Brief 008 · Education & human capability

Are people gaining the capabilities needed to adapt?

Foundational outcomes, pathway access, adult participation, teacher capacity and AI-tutor experiments answer different questions. This brief keeps them separate.

Evidence reviewed 2026-07-26World → Netherlands → selected studiesInspect sources
  1. 01Where are we now?
  2. 02What direction?
  3. 03Who gains?
  4. 04What else explains it?
  5. 05What changes our mind?

01 · Where are we now?

Capability is a sequence,
not a single score.

A learner can enter school without reaching minimum proficiency, complete a programme without finding matched work, or join training without demonstrating a new skill. No capability composite is reported.

World

Foundational learning

The 2022 learning-poverty update estimated 70% of children at late-primary age in low- and middle-income countries could not read and understand a simple text.

Europe / Netherlands

Outcomes and access

PISA 2022 recorded lower Dutch mathematics, reading and science scores than 2018, alongside large within-country gaps.

Selected studies

Intervention mechanisms

Pathways and tutoring studies identify bounded mechanisms; they do not form a country ranking.

Modelled estimate · low- and middle-income countries · 2022
70%

Estimated learning poverty

Children at late-primary age unable to read and understand a simple text. This harmonised construct combines schooling deprivation, assessment evidence and simulation; it is not a direct universal observation.

Foundations

Can learners perform?

Proficiency distributions within a named assessment.

Pathways

Can learners progress?

Enrollment, completion and labour outcomes remain separate.

Adult learning

Can adults adapt?

Participation alone cannot answer the question.

Capacity

Can systems deliver?

Staffing constructs are local and non-interchangeable.

AI tutoring

Can a task improve?

Effects need duration, comparator, costs and equity.

02 · What direction are we moving?

Measured foundations fell.
Recovery remains unproven.

Within OECD PISA—15-year-olds, the same three domains, four named cycles—Dutch mean scores fell in mathematics, reading and science between 2018 and 2022. PISA must not be joined to PIRLS, TIMSS, national assessments or adult PIAAC.

Foundational trajectory · OECD PISA

Netherlands · mathematics · age 15

Netherlands PISA mathematics scores, 2012 to 2022Mathematics mean scores were 523 in 2012, 512 in 2015, 519 in 2018 and 493 in 2022.2012201520182022523512519493
View chart as data
Netherlands PISA mathematics mean score, 15-year-old students
CycleFrameworkAgeDomainMean score
2012OECD PISA15Mathematics523
2015OECD PISA15Mathematics512
2018OECD PISA15Mathematics519
2022OECD PISA15Mathematics493

Comparability: PISA 2022 mathematics was the major domain. The Dutch exclusion rate exceeded the technical standard and may bias results upward. Test mode, cohort coverage and domain emphasis remain plausible influences.

Source: OECD, PISA 2022 Netherlands note ↗ · Scope & caveats ↓

Pathway sequence

Access is not outcome

No collapsed funnel
  1. 01Enrollmententered
  2. 02Completioncredentialed
  3. 03Employmentworking
  4. 04Earningsreturn
  5. 05Job matchskills used
Pathway stages and the distinctions required for interpretation
StageQuestionMeasureNot equivalent to
EnrollmentDid a learner enter?Participants or entrantsCompletion
CompletionWas the credential completed?Graduates or completersEmployment
EmploymentWas a recent graduate employed?Employment in a defined windowEarnings or job match
EarningsWhat return followed?Earnings with a comparison groupSkill use or job quality
Job matchAre acquired capabilities used?Field or skill matchAdaptability

Official annual context · Netherlands

Primary gross enrollment · access proxy

Enrollment ≠ outcome
View complete selected series
Netherlands primary gross enrollment ratio, annual observations
YearValueUnitEvidence
2016105.7% gross enrollment ratioObservation
2017105.0% gross enrollment ratioObservation
2018104.6% gross enrollment ratioObservation
2019104.6% gross enrollment ratioObservation
2020104.3% gross enrollment ratioObservation
2021104.4% gross enrollment ratioObservation
2022104.2% gross enrollment ratioObservation
2023104.2% gross enrollment ratioObservation

Definition: Total primary enrollment, regardless of age, divided by the official primary-school-age population. The ratio can exceed 100; enrollment is not attendance, completion or proficiency.

Source: World Bank / UNESCO UIS, SE.PRM.ENRR ↗ · Scope & caveats ↓

03 · Who gains or remains excluded?

Averages conceal
different capability gaps.

Each gap below retains its own domain, subgroup definition and denominator. The adjusted migration-background result remains an association, not a causal estimate.

Minimum proficiency73%

At or above PISA Level 2 in mathematics; all represented Dutch 15-year-old students.

Minimum proficiency65%

At or above PISA Level 2 in reading; all represented Dutch 15-year-old students.

Socioeconomic status106 pts

Mathematics gap: top versus bottom national quartile of PISA ESCS.

Migration background27 pts

Mathematics gap favouring non-immigrant students after adjustment for socioeconomic profile.

Gender26 pts

Reading gap favouring girls over boys in the represented population.

Netherlands PISA 2022 proficiency distribution and gaps
MeasureDomainValueSubgroup denominator
At or above Level 2Mathematics73%All represented 15-year-old students
At or above Level 2Reading65%All represented 15-year-old students
Advantaged–disadvantaged gapMathematics106 score pointsTop vs bottom national ESCS quartile
Adjusted non-immigrant–immigrant gapMathematics27 score pointsMigration-background groups after ESCS adjustment
Girls–boys gapReading26 score pointsRepresented girls vs represented boys

Source: OECD, PISA 2022 Netherlands note ↗ · Scope & caveats ↓

Adult learning & adaptation

Participation is the first column, not the verdict

Two windows · do not splice
EU-LFS · four weeks · 202127%

Netherlands, ages 25–64, formal or non-formal education and training. EU: 11%. A 2021 methodology break applies.

AES · twelve months · 202225.1%

EU adults aged 25–64 with low prior attainment participated. This is a different survey and reference window.

Adult-learning participation evidence; reference windows are not comparable
SourceGeography / subgroupAgeWindowValueWhat it does not show
EU-LFS 2021Netherlands25–64Previous 4 weeks27%Completion, skill gain or transition
EU-LFS 2021European Union25–64Previous 4 weeks11%Completion, skill gain or transition
AES 2022EU, low prior attainment25–64 comparisonPrevious 12 months25.1%Measured reskilling or job change

Boundary: participation ≠ completion ≠ skill gain ≠ occupational transition. The AES and LFS windows cannot be treated as a trend.

Sources: Eurostat EU-LFS ↗ + AES ↗

Teacher capacity · local construct

72% of represented Dutch students were in schools where principals said staff shortages hindered instruction.

The corresponding share was 36% in 2018; 46% in 2022 were in schools reporting inadequate or poorly qualified staff as a constraint.

Principal-reported teaching-staff constraints, Netherlands PISA
ConstructYearValueDenominator
Lack of teaching staff hinders instruction201836%Represented students, weighted by principal response
Lack of teaching staff hinders instruction202272%Represented students, weighted by principal response
Inadequate or poorly qualified staff hinder instruction202246%Represented students, weighted by principal response

Construct boundary: principal-reported constraint ≠ pupil-teacher ratio ≠ vacancy ≠ unqualified teaching ≠ workload ≠ forecast shortage. Cumulative spending of about USD 119,600 PPP per Dutch student from ages 6–15 is allocation context, not proof of outcomes.

Source: OECD, PISA 2022 Netherlands note ↗ · Scope & caveats ↓

04 · Competing explanations

Decline and gaps have
more than one mechanism.

Coverage

Cohort composition

Enrollment, exclusions and migration alter who a test represents.

Measurement

Mode and framework

Digital mode, domain emphasis and framework changes can affect comparisons.

Disruption

School closures

Closures matter, but several pre-2020 trends were already negative.

Resources

Household conditions

Language, income, devices, study space and parental time shape opportunity.

Sorting

Pathway selection

Raw graduate outcomes mix programme effects with who enters each route.

Economy

Labour demand

Employment can rise or fall without a change in training quality.

Delivery

Institutional capacity

Staffing, curriculum, procurement and timetable determine implementation.

Identification limit: timing alone does not isolate closures, household shocks, institutional capacity or cohort composition. A descriptive trend is not a causal decomposition.

05 · What would change our conclusion?

Show durable gains,
not a promising demo.

The AI studies below are deliberately not pooled. Their populations, interventions, randomisation units, comparators, durations and outcomes are incompatible.

Structured AI-tutor evidence matrix

Four experiments · four different claims

No forest plot · no pooled effect
AI tutoring and tutor-support studies; effects are not commensurable
StudyDesign & sampleComparatorDurationOutcome & uncertaintyFunding / conflictsGeneralizability
Harvard physics AI tutorRandomized crossover; 194 eligible students in one introductory courseBespoke AI tutor at home vs same content in active-learning classTwo lessons in consecutive weeksImmediate post-test: 0.63 SD linear estimate; quantile range 0.73–1.3 under ceiling effects; p<10−8No dedicated funding statement on article page; Harvard production support; authors declared no competing interestsOne selective university, course and immediate outcome; no retention, equity, cost or substitution test
Tutor CoPilotPreregistered tutor-level RCT; 783 tutors, 1,000+ K–12 studentsHuman tutors with optional AI suggestions vs tutors without themAbout 7 weeksUnconditional session exit-ticket passing +4 pp (SE 1.5 pp); no significant end-of-year MAP effectFEV Tutor and partner district collaboration; no standalone funding/conflict declarationHuman-AI support, one provider/district, low student exposure, proximal outcome
Nigeria Copilot programmeSchool-level RCT; first-year secondary students in 12 Edo State public schoolsTeacher-facilitated after-school Copilot programme vs usual schooling6 weeks0.31 SD on combined English, AI-knowledge and digital-skills assessment; study-specific standardizationWorld Bank working paper tied to Edo education operation; Microsoft product usedOne state, facilitated programme, mixed assessment; no autonomous or long-run system test
SCALE access + human supportTwo RCTs; about 350 elementary students in two unnamed U.S. districtsIndependent AI literacy access vs access plus human engagement tutor14–31 week windows; typical active use 4–5 weeksEngagement +71–80%; low usage persisted and reading achievement did not improveStanford SCALE / EdWorkingPapers; platform unnamed; no standalone funding/conflict statement in public summaryImplementation evidence, not general instructional efficacy

Boundary: AI task effect ≠ durable learning ≠ teacher substitution ≠ cost reduction ≠ system equity. No pooled estimate is defensible.

Primary sources: Kestin et al. ↗ · Tutor CoPilot ↗ · Nigeria RCT ↗ · SCALE access RCTs ↗

  1. Sustained recovery within the same assessment framework, age, domain and mode.
  2. Narrower gaps under stable subgroup definitions and denominators.
  3. Pathway evidence connecting enrollment, completion, employment, earnings and job match without collapsing them.
  4. Measured adult skill gains and occupational transitions, especially for adults with low prior attainment.
  5. Reduced vacancies, unqualified teaching and workload constraints under local measures.
  6. Independent, replicated, long-duration AI trials with retention, equity, total cost and safe system integration.
Hypothesis

AI may extend feedback and practice when teachers and institutions can integrate it.

Possibility

Work-based pathways may improve transitions when employer demand and credential quality align.

Data gap

No defensible composite joins proficiency, adult skill gain, pathway outcomes and teacher capacity.

Method & source ledger

Scope travels
with every number.

Only stable World Bank/UIS series refresh automatically. PISA, learning poverty, Eurostat survey extracts, pathway evidence, teacher constraints and AI studies remain edition-pinned and review-gated.

Updater boundary

The updater validates indicator identity, geography, definition, unit, chronology and observation count before an atomic write. A network failure preserves the reviewed file byte-for-byte; a schema or definition failure stops. Curated evidence is never rewritten.

PISA 2022 Results, Netherlands

OECD · observation · 15-year-old students · 2012–2022 cycles

Open source ↗

State of Global Learning Poverty: 2022 Update

World Bank et al. · modelled estimate · late-primary age

Open source ↗

Adult participation in learning (trng_lfse_01)

Eurostat EU-LFS · survey observation · four-week window

Open source ↗

Adult Education Survey 2022

Eurostat · survey observation · twelve-month window

Open source ↗

Employment of recent graduates (edat_lfse_24)

Eurostat · observation · ages 20–34 · 2025

Open source ↗

European Education Area policy context

European Commission / Eurostat · policy target

Open source ↗

Primary gross enrollment (SE.PRM.ENRR)

World Bank / UNESCO UIS · official annual observation · Netherlands

Open API ↗

AI tutoring outperforms in-class active learning

Scientific Reports · randomized crossover experiment · 2025

Open study ↗

Tutor CoPilot

Stanford SCALE / EdWorkingPapers · preregistered RCT

Open paper ↗

From Chalkboards to Chatbots

World Bank · school-level randomized experiment · 2025

Open paper ↗

Access is Not Enough

Stanford SCALE / EdWorkingPapers · two RCTs · 2026

Open study ↗