Two separate questions for every subject
At work: how much of the paid work will AI do? Read across any row and the answer only goes up. In school: should pupils still learn it? For all 14 subjects the answer is yes, though for some the reason changes. Click any subject to open its deep dive.
AI's share of the work never falls in any scenario. What holds steady is the case for learning the skill, because people still need it to check AI, to live and to take part as citizens.
- Small: AI assists at the edges
- Shared: AI does parts, people lead
- Large: AI does much of the routine paid work
- Most: AI does most of the paid work
- Confidence. Falls to very low by 2056
- StrongerYes, focus shiftsYes, reason changes
~2036 is anchored in evidence and official projections. ~2046 and ~2056 are scenarios, not forecasts.
| Subject | At work: how much does AI do? | In school: still worth learning? | ||
|---|---|---|---|---|
| ~2036 evidence-anchored |
~2046 scenario |
~2056 scenario |
||
Drawn from the report's per-subject outlooks. Where the report gives no separate 2056 work estimate, the cell holds the 2046 level, since no scenario has AI's share falling. Scenarios assume AI capability somewhere between the 2047 and 2116 expert medians: a faster path pushes more rows to "Large" or "Most" sooner, slow adoption holds them at "Shared".
What pupils learn, what AI threatens, what it offers
Every point links to its source. Tags show what kind of evidence it is, and a CHECK tag marks figures that were not re-checked against the primary source.
How an AI tool is designed decides whether pupils gain or lose
The strongest study so far gave about 1,000 Turkish secondary students access to GPT-4 for maths practice, then took it away for the exam.
Change in score versus a control group with no AI. Source: Bastani et al., PNAS 2025. Field experiment, about 1,000 students.
Harvard physics: a carefully scaffolded GPT-4 tutor beat in-class active learning, in less time. Only 194 students and two topics; a re-analysis suggests a smaller gain. Kestin et al. 2025EMP
Nigeria: six weeks of teacher-led GPT-4 sessions improved English. Working paper; control group had no extra lesson time. World BankEMP
England: ChatGPT with a guide cut KS3 science lesson-planning time with no drop in quality. 259 teachers. Measures workload, not pupil learning. EEFEMP
"Cognitive offloading": surveys link heavy AI use to less critical thinking, and the MIT EEG preprint (54 people) found weaker engagement. Correlational or unreviewed. Gerlich · MIT preprint
of UK 8–17-year-olds have used AI. Of those who read AI search summaries, 46% think they are always accurate. Ofcom 2026
AI malpractice cases detected at GCSE and A level in 2025: only 2% of all malpractice. Treat it as a floor. Ofqual has demanded stricter coursework checks. Ofqual
disadvantaged pupils targeted by DfE AI tutoring tools, due by end of 2027. No English pupil-facing AI trial has reported results yet. GOV.UK
Exposure is broad. Actual job loss so far is narrow, and it falls on junior roles.
Exposure studies estimate what AI could touch. They are not job-loss forecasts. The DfE's own study says it "does not distinguish whether a job will be augmented or replaced".
US 22–25-year-olds in the most exposed jobs, against trend. Descriptive, not causal. StanfordEMP
Junior roles at highly AI-exposed UK firms. Customer-facing roles rose slightly. King's College LondonEMP
Senior vs junior hiring at AI-adopting firms in 41 countries. AI changes the mix more than the total. Not yet peer-reviewed. Phys.org
fewer US translator jobs created 2010–23 as machine translation spread. The clearest causal displacement evidence of any subject. Frey & Llanos-ParedesEMP
Earnings of image freelancers on Upwork after AI image tools (writing: −5.2%). Hui, Reshef & ZhouEMP
Speed of experienced developers on their own code when using AI, though they believed it helped. METR 2025EMP
- Management consultants, accountants, psychologists
- Teaching
- Jobs reached through maths and economics degrees
- London and the South East
- Sports players (the least exposed of all)
- Roofers, steel erectors and other trades
- Hands-on care and science roles: vets, radiographers, physios
Source: DfE, The impact of AI on UK jobs and training (2023). Generative AI reverses the old automation pattern: the more educated and better paid are more exposed.
| Headline estimate | What it measures | Read it as |
|---|---|---|
| IPPR: 11% of UK tasks | Tasks exposed to current AI; 59% in a "second wave" | Scenarios from 7.9m UK jobs lost to none, depending on policy. IPPRMODEL |
| NFER: 1–3m UK jobs by 2035 | Declining admin, secretarial, customer-service and machine roles | Total employment still grows, in roles using six essential skills. NFERMODEL |
| McKinsey: 57% of US work hours | Technical potential of today's technology | "Not a forecast of job losses." MGI 2025VENDOR |
| Acemoglu: ~4.6% of US tasks | Tasks profitably automatable within ten years | The sceptical end. NBERCHECK |
| Frey & Osborne 2013: 47% | US jobs at "high risk" | Never materialised as job loss; a task-based re-analysis cut it to 9%. OECD 2016CHECK |
The safest bets hold up in every scenario
No credible forecast covers specific school skills beyond 2035. So the sound approach is to invest in skills that stay valuable whichever future arrives.
The foundation for spotting AI errors. 21% of adults in England are at or below Level 1 numeracy.
Lateral reading improved pupils' credibility judgements in six lessons. Analytical thinking is the WEF's top skill.
The strongest long-run evidence of rising value (Deming 2017). Customer-facing roles held up in UK data.
Least exposed to generative AI in every study. Robotics is the open question at 20–30 years.
The CAR wants pupils to use AI "without becoming dependent on it". DfE now asks tools to monitor cognitive offloading.
| Skill family | Evidence now | Next 10 years | 20–30 years (speculative) | Evidence base |
|---|
Sources: WEF Future of Jobs 2025 (employer survey: reading, writing and maths are among the most stable skills, not obsolete) · NFER Skills Imperative 2035 (essential skills almost unchanged since 2010) · Deming, QJE 2017 · KCL · DfE 2023.
A child starting Reception in 2026 sits GCSEs around 2037 and reaches mid-career around 2056
Hard evidence covers about 2022 to 2026. Official projections reach about ten years. Everything later is a scenario.
Evidence tags, and what to check before quoting
- UK graduate vacancies down 45.6% (Adzuna via a secondary source)
- 2026 FrontierMath and SWE-bench scores from unofficial aggregators
- Deezer's share of AI-generated uploads (company-reported)
- The 32% illustrators figure (secondary source, sample unknown)
- AllFaith religious-bias results (press coverage, partly preprint)
- Exact wording of DfE's "cognitive offloading" standard
- OECD Digital Education Outlook 2026 (secondary summaries only)
- Some programme-of-study wording (English, history, geography) not re-fetched
- Whether the revised RSHE guidance names companion chatbots
- No UK school trial of AI used directly by pupils has reported yet
- No study tracks skills over more than about two years of AI use
- No forecast covers any school subject's skills at 20 or 30 years
- No UK payroll-data equivalent of the Stanford analysis
Scope: the national curriculum applies to England. Scotland (Curriculum for Excellence), Wales (Curriculum for Wales) and Northern Ireland have their own curricula and mostly treat AI as a cross-curricular safeguarding and digital-competence issue.