AI Must Not Remove the First Rung of Work

An institution can improve today's output while weakening tomorrow's capability. The danger appears when artificial intelligence absorbs the elementary tasks through which a beginner learns how work is actually done.

Entry-level work is often repetitive. It can include first drafts, routine research, simple coding, document preparation, meeting notes, basic analysis and supervised contact with clients, learners or communities. Those tasks may look like obvious candidates for automation. Yet they also teach standards, vocabulary, exceptions, relationships and the consequences of error. They place the novice close enough to experienced people to observe judgment and receive correction.

The governing question is therefore not only whether AI can complete a junior task. It is whether the organisation can remove that task without removing the pathway by which somebody becomes capable of senior responsibility.

Responsible adoption should preserve a visible first rung: paid entry → supervised practice → feedback → progressively independent performance → accountable judgment → responsibility for forming another person.

The newest warning concerns the beginning of professional life

On 12 October, Singapore's President Tharman Shanmugaratnam warned at the annual meetings of the International Monetary Fund and World Bank that AI may reduce demand for fresh graduates in the sectors most exposed to it. He described the first job after graduation as the point at which demand is already becoming weaker and also warned that AI could deepen divergence between advanced and developing economies. Reuters, 12 October

This is a consequential warning from a national leader with longstanding economic-policy experience. It is not a causal labour-market study, and it does not prove that AI has produced a general fall in graduate employment. Interest rates, post-pandemic hiring corrections, sectoral cycles, changing work locations and employer expectations may also affect the same jobs.

Its importance lies in the distributional question it raises. Even when total employment remains stable, the route into a profession may narrow. A labour market can continue to employ experienced people while asking where the next generation of experienced people will come from.

The employment evidence is mixed and should remain so in the account

A September policy brief from the Stanford Institute for Economic Policy Research concludes that aggregate AI effects on employment are probably still small. It notes evidence of declining employment among younger workers in highly AI-exposed occupations, but also finds that unemployment in those occupations has not generally risen faster. The brief emphasises how difficult it is to separate AI from interest rates, technology-sector over-hiring during the pandemic and the decline of remote work. Stanford SIEPR

A UCLA working paper reported in late September found no statistically significant AI-specific increase in unemployment among recent United States college graduates during the summer of 2026. The unemployment rate for graduates aged twenty-two to twenty-five was 7.3 per cent, higher than a year earlier but within normal variation. Washington Post summary

The working paper is early evidence from one national labour market, not a final answer. Occupational exposure does not reveal how individual employers reorganise work, and unemployment alone may miss fewer vacancies, longer searches, lower-quality entry roles, reduced hours or the replacement of paid employment with unpaid tests and platform tasks.

The defensible conclusion is narrower than either optimism or alarm: a universal AI-driven graduate-employment collapse has not been established, but credible sectoral and early-career risks justify measurement and action before the pathway is lost.

Entry-level work is part of a profession's learning infrastructure

A junior assignment has at least two outputs. It produces the immediate document, analysis, lesson, visit, programme or piece of code. It also produces a more capable worker. Conventional productivity accounting usually values the first output and hides the second.

Automation can therefore appear beneficial when it shortens the task while transferring verification to a senior employee and removing the novice's opportunity to learn. The organisation records time saved but not supervision displaced, errors missed, professional identity weakened or future recruitment made more difficult.

This is not an argument for preserving pointless drudgery. Poorly designed entry work can be exploitative, repetitive and detached from progression. The proper response is to distinguish unnecessary burden from formative practice. Automate what adds no developmental value, redesign what can become deeper practice and protect the experiences through which judgment grows.

Every profession should be able to answer: Which early tasks reveal how the system works? Which expose novices to real people and consequences? Which build memory, pattern recognition and ethical responsibility? Which can AI assist without replacing the learner's own performance?

Productivity gains do not decide how the gains should be distributed

Studies and field experiments frequently find that generative AI can improve performance on bounded tasks, often with particularly large gains for less experienced or lower-performing workers. That finding supports assistance, but it does not settle the organisational design.

The same productivity gain can produce different futures. An employer can use it to remove junior roles, intensify workloads or concentrate rewards among owners and experienced staff. It can also shorten low-value work while keeping paid entry, improving coaching and giving novices access to more varied and consequential assignments.

The choice is governed, not technologically predetermined. Boards and executives determine headcount, workload, mentoring time, pay, procurement and the evidence by which a pilot is judged. A business case that counts licences and hours but excludes recruitment, succession, capability and distribution is incomplete.

An institution should state who receives the saved time, who loses paid opportunity, which capability must still be formed and how the productivity benefit will support human development. Efficiency becomes public or institutional value only through such choices.

Curriculum should prepare learners to work with, without and about AI

Education cannot promise that graduates will enter the same tasks their predecessors performed. It can ensure that they understand the foundations beneath automated work, use assistance responsibly and evaluate whether a system should be trusted or used at all.

A resilient curriculum develops three connected capabilities. Learners should perform essential responsibilities without technological substitution. They should use approved tools under supervision, verifying evidence and disclosing material assistance. They should also reason about data, incentives, limitations, labour effects, professional duties and the option to refuse unsafe automation.

The sequence matters. A learner who delegates the first encounter with a problem may receive fluent output without building the structure needed to judge it. A learner who never uses AI may lack the supervision skills now expected in practice. Curriculum should move deliberately from independent foundations to bounded assistance and then to accountable professional integration.

A useful sequence is: encounter → attempt → receive feedback → use AI to challenge or extend → verify → revise with reasons → transfer to an unfamiliar task → perform independently.

Assessment must reveal whether capability survives the tool

A polished submission cannot show who performed the reasoning. Assessment needs complementary evidence: an initial independent response, process notes, source verification, version history, oral defence, an unfamiliar application, observed practice and a later task completed with reduced assistance.

The purpose is not surveillance or the impossible detection of every AI-generated sentence. It is to make learning and responsibility visible. A learner should be able to explain the decision, locate the evidence, identify uncertainty, correct an error and act when the system is unavailable or wrong.

Assessment can also examine AI supervision directly. Give learners a flawed output, an ambiguous instruction or a culturally inappropriate recommendation. Ask them to diagnose the failure, protect affected people, revise the workflow and justify whether use should continue.

The institution should measure progression rather than one-time fluency. Is the learner becoming less dependent on prompts? Can knowledge transfer to a new context? Can the learner recognise when human presence, consent or professional authority is required? These are stronger indicators of employability than rapid completion alone.

E-learning should make the apprenticeship ladder visible

A learning-management system can support progression when it connects preparation, supervised practice, feedback and independent performance. It can require an initial response before revealing examples, schedule mentor observation, retain a record of revision and provide a later transfer task.

It can also flatten formation into content consumption and automated completion. Platform time, chatbot turns, badges and submission speed do not demonstrate judgment, character or professional readiness. An AI tutor that answers continuously may conceal the absence of a lecturer, mentor or workplace opportunity.

Digital design should therefore identify the human relationship required at each stage. Who observes the learner? Who can correct a misunderstanding? Who decides that competence is sufficient for greater responsibility? How does a learner appeal an unfair judgement or obtain an accessible alternative?

Low-bandwidth and asynchronous routes remain essential. If the approved pathway requires an expensive device, continuous connectivity or a paid AI subscription, the institution may be measuring infrastructure privilege rather than capability.

Employers should redesign entry roles before eliminating them

A responsible employer begins with task and pathway mapping. Identify which junior activities AI can accelerate, which experiences build tacit knowledge, which relationships give the work meaning and which risks require qualified human authority.

Then redesign the role. A junior analyst may spend less time assembling standard tables and more time checking source quality, comparing interpretations and presenting uncertainties. A beginning teacher may use AI to generate alternatives after designing an initial lesson, then observe learners, adapt instruction and defend the final choices with a mentor. A junior developer may use code assistance while remaining responsible for tests, security reasoning and explanation.

Protected mentoring time is a production input, not optional generosity. Senior employees should have recognised workload for demonstration, feedback and review. Novices need progressive access to real work rather than permanent confinement to synthetic exercises or the verification of machine output.

Employers should publish the pathway: entry criteria, paid learning, permitted assistance, evidence of progression, decision authority and the conditions for greater responsibility. Transparency helps graduates prepare and allows boards to see whether succession is real.

Global South workers should not carry the cost of transition

The first rung of digital work is already uneven. Data labelling, transcription, content moderation, translation and platform tasks have created income for workers whose contribution is often hidden behind products marketed elsewhere. Automation may reduce some of these opportunities while demand moves toward specialised evaluation, safety testing and domain expertise.

A transition that removes low-paid work without creating accessible progression is not inclusion. Workers need transparent terms, fair pay, occupational safety, portable evidence of skill and routes into higher-value responsibilities. Training should not be offered as a substitute for income when no realistic job pathway exists.

An evaluation of career-guidance systems across ten African countries found weak contextual awareness of local education, infrastructure and culture. The study is a preprint and its model sample will not represent every product, but it illustrates why imported advice should not direct young people's futures without local verification. African career-guidance evaluation

The African Union's Continental Artificial Intelligence Strategy appropriately connects adoption with African agency, infrastructure, skills, culture, inclusion and responsible governance. Agency means helping Africans shape and own the higher-value parts of the AI economy—not only supplying data, consuming tools or absorbing displaced risk. African Union

Theological education must preserve supervised ministry practice

Ministers are not formed by text production alone. They learn by handling Scripture, listening to people, preparing and delivering teaching, visiting homes and hospitals, receiving correction, responding to conflict and serving under accountable leadership.

AI may support research, translation, rehearsal, administration and the generation of questions. It cannot carry the learner's character, calling, responsibility or relationship with a congregation. Removing elementary ministry assignments because software is faster would confuse output with formation.

Competency-based theological education integrates knowing, being and doing in context. Mentors and ministry communities should witness performance over time, including how a learner adapts when a prepared response no longer fits the person or situation. CBTE framework

A formation pathway can use technology without surrendering apprenticeship: observe faithful practice; assist with a bounded responsibility; prepare independently; use AI for challenge where permitted; verify theology and evidence; perform under observation; receive correction; perform in a new context; and eventually mentor another learner.

Digital ministry should reinvest saved time in people

Churches may use AI to organise public information, prepare routine communication, improve accessibility or reduce administrative load. The ethical question is what happens to the time released.

If efficiency simply reduces human contact or concentrates visible ministry among a few senior leaders, the church may become more productive in media while becoming weaker in discipleship. Saved time should return to people through prayer, teaching, visitation, mentoring, safeguarding and the careful formation of emerging leaders.

Young ministers need genuine, bounded responsibility. They should not be used as unpaid labour, placed alone in high-risk pastoral situations or given platform visibility without character formation. They need clear authority, supervision, feedback, protection and a route to greater responsibility.

AI must not become the source of prophecy, spiritual diagnosis or pastoral authority. Generated religious fluency is not anointing, and automated availability is not accountable presence. Technology may assist the work; it cannot become the minister or remove the relationships through which ministers are formed.

Boards need a First-Rung Impact Assessment

Before approving an AI system or workflow redesign, a board should require a First-Rung Impact Assessment. It should record the problem being addressed, the current workflow, tasks to be automated, people affected, entry opportunities changed, capabilities formed by those tasks and the redesigned pathway to independent responsibility.

The assessment should also identify mentoring capacity, accessibility, data and permissions, expected productivity, full costs, distribution of gains, potential displacement, appeal, review and exit. It should distinguish a temporary pilot from a permanent workforce decision.

Useful measures include the number and quality of paid entry opportunities, time to competent independent performance, error detection, source verification, mentor workload, retention, progression, pay, unequal effects, worker voice and the institution's ability to fill senior roles internally.

Aggregate headcount is insufficient. An organisation may report stable employment while replacing a developmental role with insecure task work, reducing access for people without elite networks or leaving juniors permanently responsible for checking output without understanding the profession beneath it.

Research should compare pathways, not only productivity

A recent qualitative preprint based on fourteen interviews in South Korean software engineering argues that generative AI may move routine entry work into senior-plus-AI workflows and reduce the productive struggle through which junior engineers develop. Its small, sector-specific sample cannot establish general labour-market effects, but it identifies a mechanism that larger studies should test. Kim and colleagues

A useful research model is: AI capability → task redesign → entry opportunity → supervised practice → independent performance → progression → organisational capability and renewal.

Mentorship, feedback quality, assessment design, AI literacy, pay and worker participation may mediate the pathway. Language, gender, disability, connectivity, rurality, institutional size, labour protection and market concentration may change the effects.

Comparative studies should examine conventional apprenticeship, unstructured AI use, AI adoption that reduces junior roles and a formation-protected model. Outcomes should include hiring, pay, search duration, time to competence, unsupported performance, mentor workload, retention, mobility, unequal effects and succession—not output speed alone.

Christian stewardship forms people who can carry responsibility

Christian leadership cannot treat people merely as costs attached to output. Work carries dignity, provision, service and participation in a community's shared life. Formation requires more than access to information; it involves imitation, correction, trust and progressively accountable responsibility.

Jesus called disciples to be with him before sending them. Paul entrusted work to Timothy and Titus while continuing to instruct them. The pattern is neither passive dependency nor instant independence. Mature leadership reproduces capable and faithful people.

Pentecostal communities rightly celebrate empowerment, gifting and multiplication. Those claims become credible when experienced leaders make room for emerging ones, test gifts within accountable community and value character and fruit above fluent performance.

An institution may legitimately automate a task. It must not automate away its responsibility to form the people who will teach, govern, serve and discern after today's leaders are gone.

Audit the first rung before automating it

Select one AI-enabled workflow and map the present route from novice to trusted responsibility. Identify which entry tasks create real value, which merely create burden and which apparently simple experiences carry hidden learning.

Keep or redesign the formative work. Specify the learner's initial independent attempt, AI's permitted role, the responsible mentor, evidence of verification, the next unfamiliar performance and the point at which assistance will be reduced. Protect sensitive data and preserve human authority over consequential decisions.

Then follow the people, not only the output. Did paid access become narrower? Did senior review increase? Did the novice become more capable? Were gains shared? Can the institution still grow its own future leaders?

AI should help people climb the ladder of competence. It must not remove the first rung and then blame the young for failing to rise.