The best path is neither “everyone must retrain” nor “replace wages with one universal payment.” It is to make economic security less dependent on employment, while sharing the benefits of AI through better services, broader ownership, shorter working time, and supported transitions into useful work.
That approach should expand as evidence of displacement accumulates—not wait for mass unemployment, but not assume it is inevitable either.
The central distinction is between producing prosperity and distributing access to it. AI could make society richer while making many workers poorer. It could also allow people to work less, receive better services, and spend more time on relationships, learning, and community. Technical capability alone does not determine which outcome occurs.
1. What the evidence establishes—and what remains uncertain
AI can improve productivity without immediately reducing employment
A major customer-support study found that AI assistance increased issues resolved per hour by approximately 15%, with larger benefits for less-experienced workers. That demonstrates useful productivity gains in a particular setting, not an economy-wide employment outcome.
Research linking Danish adoption surveys to administrative records found substantial changes in tasks but no detectable average effects exceeding 2% on earnings or recorded hours during the first two years after ChatGPT’s launch. The March 2026 revision emphasizes that organizations can absorb early AI adoption through work reorganization before aggregate employment or pay changes become visible.
However, entry-level opportunities deserve particular attention. An August 2026 Stanford study using ADP payroll data through June found no widespread economy-wide displacement in its sample, but employment among 22–25-year-olds in highly exposed occupations stood
19% below the level implied by keeping pace with less-exposed peers
. Reduced hiring, rather than increased separations, was the main channel. The authors explicitly describe this as observational evidence, not a causal estimate of AI’s effect; pre-existing trends and sample differences matter.
These findings are compatible: AI may help an individual junior employee perform better while reducing how many junior employees firms recruit.
Exposure is not the same as job destruction
The ILO’s 2025 assessment places approximately one-quarter of global employment in occupations with some generative-AI exposure. Clerical work is particularly exposed, alongside increasingly digitized professional work. But the assessment concerns task exposure under contemporary capabilities—not the number of jobs that would disappear under much stronger future AI.
The relevant sequence is:
Technical capability → reliable and economical deployment → organizational adoption → changes in required human hours → changes in employment, wages, and working time.
Each step introduces uncertainty. Demand might expand when services become cheaper. New businesses might emerge. Regulation, liability, physical infrastructure, and organizational change might slow deployment. Conversely, AI might automate new tasks almost as quickly as they appear.
My assessment: there is strong reason to prepare for substantial restructuring, but insufficient evidence to attach a credible date or percentage to “most jobs disappear.” Policy should work under both modest disruption and severe displacement.2. Four futures: what the arithmetic reveals
Model structure
Consider a hypothetical economy at the same future date—illustratively 2036—with a no-additional-AI baseline of:
100 employed people working 40 hours weekly.
100 units of real national income, after production costs.
60 units paid as labor income and 40 as non-labor income.
These are illustrative assumptions, not observations about any country.The assumed labor-income shares are respectively 58%, 45%, 30%, and 40%, versus 60% in the baseline. The strongest scenario assumes substantial spillovers beyond narrow knowledge tasks.
These are employment slots in a simplified fixed cohort, not predicted unemployment rates. Work-sharing is assumed frictionless, so its employment results are optimistic upper bounds.
Three conclusions follow.
First, a richer economy need not produce a larger wage bill. In the fast scenario, total income rises 30%, yet labor income falls slightly. Whether ordinary households become better off depends on ownership, transfers, taxes, prices, and services—not productivity alone.
Second, shorter hours can distribute remaining work, but cannot automatically preserve pay. A 32.5-hour week spreads work across everyone in the fast scenario, but total labor earnings remain 2.5% below baseline unless another mechanism changes the distribution.
Third, the hardest transition may involve weak growth rather than spectacular abundance. Jobs can disappear in particular activities before cheaper intelligence translates into affordable housing, functioning infrastructure, new investment, or stronger demand elsewhere.
The cash guarantee fills covered workers’ earnings to 60% of the old average gross wage. It is not a household poverty-line guarantee. The combined package adds a services allowance equal to 2% of national income; its potential employment and productivity benefits are not assumed.
In this case, work-sharing spreads the existing wage bill across more people, reducing the cash requirement. But 25 people still lack paid-work slots. A four-day week is not a complete solution to very large reductions in labor demand.
Testing the financing
Assume government captures 30% of additional non-labor income and receives 25% of changes in labor income through taxes, including lost receipts when wages fall.
In the very-fast case, incremental revenue equals 10.5% of modeled income, enough to cover the 8% package. At only 10% capture of additional non-labor income, incremental revenue falls to 1.8%: the same package becomes substantially underfunded.
In the bottlenecked scenario, the package costs 8.2%, but incremental revenue is only 2.3%.
Therefore, “tax the AI windfall” is not a sufficient financing strategy. The windfall may be small, concentrated abroad, difficult to identify, or difficult to tax.
I also varied capture rates from 10–50% and minimum workweeks from 28–36 hours. The recurring lesson is that work-sharing, income protection, and financing address different problems. None substitutes for the others.
The model does not estimate behavioral responses, prices, investment, tax avoidance, household inequality, or transition dynamics. It does not establish that the combined package maximizes welfare. It establishes conditions under which particular promises are arithmetically feasible.
3. Who is most at risk?
A useful risk assessment combines task exposure, speed of change, financial vulnerability, and ability to move into other work.
The gender dimension is important without being universal: the ILO estimates that, in high-income countries, its highest-exposure category accounts for 9.6% of female employment versus 3.5% of male employment. That reflects occupational patterns, not inherent differences in adaptability.
Experienced professionals may initially be protected by relationships, institutional knowledge, or responsibility. That is not permanent immunity. Conversely, less-exposed workers can be affected indirectly if displaced professionals compete for their jobs or local spending declines.
Support should not require workers to prove that AI caused their hardship. That creates an unnecessary attribution problem and excludes people whose opportunities vanish through reduced hiring rather than dismissal.
4. Which solutions survive contact with evidence?
Retraining works best when it leads to a real job
A randomized evaluation of Year Up—an intensive program combining training, support, and internships—found roughly 28% higher quarterly earnings at its six-year confirmatory endpoint. That supports employer-connected transitions, not the proposition that inexpensive online courses solve displacement.
Long-run WorkAdvance results illustrate the uncertainty. At year ten, one provider increased average earnings by 32%; the other three had no impact on the prespecified year-ten outcomes, although all had generated earnings gains at some earlier point.
Recommendation: fund pathways with credible employer demand, paid practice, and measured earnings outcomes. Do not subsidize training indefinitely for occupations without sufficient vacancies.
Retraining can improve a person’s position in the hiring queue. It cannot, by itself, ensure that the queue contains enough jobs.
Cash support is useful, but not a complete social system
In a US randomized experiment, 1,000 low-income adults received $1,000 monthly for three years. Recipients reduced paid work by approximately one to two hours per week. The experiment does not establish what a permanent, nationwide, tax-funded basic income would do to wages, prices, or investment.
Finland’s basic-income experiment found small employment effects and better reported well-being among recipients, with interpretation complicated by other policy changes and survey limitations.
Recommendation: establish reliable minimum-income protection now; expand unconditional dividends as durable financing becomes available. Do not replace disability, housing, and other high-need support with an inadequate flat payment.
Shorter working time deserves serious trials
A 2025 multinational study of four-day-workweek interventions found improvements in burnout, job satisfaction, and health. Participating organizations were selected, however; it does not prove that every industry can maintain output and pay while reducing hours.
Recommendation: test negotiated reductions sector by sector. Measure service quality, workload intensity, weekly pay, staffing, and output—not merely employee enthusiasm.
Broad ownership matters, but funds take time to build
A social wealth fund can eventually distribute investment income. It is not an immediate source of free money.
For illustration, in an economy producing $60,000 per adult annually:
A $12,000 universal annual payment costs 20% of GDP gross.
A $3,000 dividend costs 5% of GDP gross.
Financing that smaller dividend from a fund paying out 3% annually requires assets equal to approximately 167% of GDP.
Those are arithmetic examples, before tax recovery, benefit interactions, or administration—not country estimates.
Recommendation: begin building broad ownership early, while financing current protection through current revenues and credible borrowing capacity. Do not count the same public investment returns twice.
5. What governments should do
A. Guarantee continuity of income and essential services
The immediate objective should be preventing a job shock from becoming a housing, health, education, and family crisis.
A proposed starting design is capped earnings replacement of roughly 60–75% for six to twelve months, followed by a durable minimum-income floor. These are policy parameters to cost locally, not universally optimal rates. New entrants without an earnings history need a separate floor and transition allowance.
Benefits should remain available across employee, contractor, and self-employed status. Health coverage and basic pension protection should not disappear with a job.
Use automatic enrollment where possible, assisted and offline access where necessary, and a smooth benefit phaseout. Taking additional paid work should increase disposable income after all interacting taxes and benefit withdrawals.
B. Create a funded transition pathway, not just a training voucher
A displaced worker should receive a caseworker-supported sequence:
Rapid income stabilization → assessment of transferable skills → a choice among viable pathways → paid training or placement → follow-up after reemployment.
Include childcare, transport, accessibility accommodations, and recognition of prior experience. Training should be modular so people do not repeatedly start from zero.
Wage insurance can help when a worker accepts a useful but lower-paid role. One proposed design would temporarily replace half the earnings loss, subject to a cap, for up to two years. Research on a US trade-displacement program supports the potential of wage insurance to improve reemployment and cumulative earnings, but does not establish universal self-financing under an AI shock.
Prevent abuse through minimum labor standards, employer contributions, and checks against using public subsidies to replace existing workers with cheaper ones.
C. Share productivity gains through both pay and time
Support collective negotiation, profit-sharing, and temporary work-sharing arrangements rather than treating layoffs as the default adjustment.
Where verified productivity gains permit, reduce weekly hours while protecting low and middle incomes. Where they do not, acknowledge the cost instead of promising unchanged pay by decree.
Public support should be conditional on demonstrable sharing of benefits and should not permanently sustain firms without viable demand.
A voluntary public-service employment option can provide paid work in environmental restoration, community support, accessibility, and other locally identified needs. It should offer useful work, proper supervision, and genuine choice—not punitive workfare or invented tasks performed merely to qualify for subsistence.
D. Broaden ownership and modernize the revenue base
My recommended direction is to reduce excessive dependence on taxing payroll and strengthen taxation of capital income, economic rents, inheritances, and immovable land, adapted to local institutions.
This is preferable to a blunt tax on every AI installation, which would also penalize socially valuable adoption. The IMF’s fiscal analysis similarly emphasizes stronger social protection and capital-income taxation rather than a general “robot tax.”
Practical ownership measures include diversified employee savings and ownership arrangements, public investment funds, and transparently priced public equity or warrants in return for substantial public subsidies.
Workers should not have both their job and their entire financial future tied to one employer. Social funds should have independent governance, transparent mandates, and protections against political patronage.
International tax coordination matters because the jurisdiction losing employment may not be the jurisdiction receiving the profits. The IMF’s 2026 scenario exercise identifies state capacity, diffusion, and cross-border capture of gains as major determinants of outcomes.
E. Make essentials cheaper—not just intelligence
Transfers buy little security when housing, energy, care, and transport remain inaccessible.
Pair redistribution with housing construction, grid investment, efficient public procurement, preventive health services, and expanded care capacity. AI should help reduce administrative costs and waiting times, but savings should be measured rather than presumed.
Competition policy is also central. OECD analysis identifies concentration, switching barriers, and infrastructure bottlenecks in AI markets. My recommendation is interoperable procurement, portability, scrutiny of exclusionary arrangements, and practical access for smaller firms and public institutions.
Avoid a two-tier settlement in which affluent people receive accountable human services while everyone else is forced into inferior automated systems.
F. Adapt the package to poorer countries
A rich-country dividend model cannot simply be copied where fiscal capacity and infrastructure are limited.
Priorities should include reliable electricity, connectivity, local-language tools, basic payment and identification systems with offline alternatives, and applications in agriculture, health, and education. The World Bank emphasizes connectivity, compute, contextual data, and competencies as foundations for inclusive adoption.
Development finance should support these foundations and diversification in exposed service-export economies. Countries do not all need their own frontier model; they need affordable, dependable access and enough local capability to use it on local problems.
6. What people should learn—and which work may remain valuable
There is no intellectually honest list of permanently “AI-proof” knowledge jobs under the premise that AI can perform most knowledge work.
“Become a prompt engineer” or “supervise the AI” is not a sufficient long-term answer. Those activities can themselves become automated.
Nevertheless, some kinds of work have additional reasons for human participation. The following are conditional opportunities, not guaranteed employment forecasts.
Type of workWhy human participation may remain valuableImportant caveatPhysical installation, repair, and field operationsWork occurs in varied real-world environments and requires physical executionRobotics can change the boundary; training and physical suitability matterCare, education, coaching, and mediationRelationships, trust, consent, and continuity can be part of the service itselfSocial value does not automatically produce adequate wagesAccountable implementationOrganizations need someone authorized to make commitments and answer for outcomesAccountability can concentrate in fewer people rather than preserve every roleScientific and industrial deploymentTurning an idea into a tested treatment, product, building, or process requires real-world executionAI may accelerate—and also automate—parts of this workHuman-centered experiences and community workPeople may value human performance, companionship, participation, and authenticityWillingness to pay and public funding constrain employment
The better career question is not “Which title is safe?” It is:
What useful outcome can I help deliver, who needs it, why would they pay for my participation, and how might that answer change?
What companies should do
Companies should assess AI against end-to-end outcomes: quality-adjusted output, errors, security incidents, customer outcomes, and the cost of review. Tool usage and generated content are not productivity measures.
Before major displacement, require an internal transition plan covering paid learning time, redeployment, consultation, and support for employees who cannot move into the remaining roles. Retain meaningful junior pathways rather than assuming an experienced workforce can reproduce itself without apprenticeship.
Share verified gains through some combination of lower prices, better pay, reduced hours, employee ownership, and investment. Publish enough information for workers to understand whether “AI transformation” means improvement or simply intensified workloads.
Do not make employees individually responsible for financing the transition while shareholders receive the entire upside.
What schools should teach
Education should develop both independent competence and capable use of tools.
Keep writing, mathematics, statistics, scientific reasoning, history, domain knowledge, and the ability to explain an argument without AI. Add verification, uncertainty assessment, data protection, workflow design, and understanding of institutional responsibility.
A mathematics experiment found that unguarded generative-AI assistance could improve practice performance while damaging subsequent unaided performance; tutoring safeguards mitigated that problem.
Conversely, a six-week teacher-guided AI tutoring program in Nigeria improved assessed learning. Because extra instruction, teachers, and AI were bundled, the result is evidence for a supported educational program—not proof that teachers can be removed.
Use a mixture of AI-assisted projects, oral defenses, supervised unaided assessment, and real-world work. Educate people for citizenship, relationships, and a meaningful life as well as employment.
What individuals should do over the next 90 days
Start by breaking your work into tasks: what can be automated, what benefits from assistance, and what depends on relationships, physical execution, or responsibility.
Then build one useful AI-assisted workflow and measure the result—including mistakes and review time. In parallel, explore two adjacent career paths through employer conversations, small projects, or paid placements before purchasing expensive retraining.
Learn a combination of domain expertise, verification, quantitative reasoning, communication, and practical implementation. Build evidence that you can solve a real problem, not merely operate a fashionable interface.
Maintain relationships and activities outside employment. Financial buffers help where feasible, but resilience advice must not become a way of blaming people who lack the income to save.
7. Relationships, mental health, and purpose need their own plan
Income protection is necessary, but insufficient.
WHO identifies good work as a source of income, routine, social inclusion, and purpose, while insecurity, excessive demands, and poor working conditions can harm mental health. The relevant goal is therefore not simply to maximize employment at any cost.
Two people can have identical paid hours but very different lives. One chooses a shorter week with secure income and meaningful activities. Another is excluded from employment, financially anxious, and ashamed. Calling both situations “less work” obscures the difference.
A good transition should protect agency, belonging, contribution, and recognition.
Governments and communities should invest in libraries, sports, arts, open workshops, adult education, volunteering infrastructure, and accessible shared spaces. People need opportunities to make commitments, develop mastery, help others, and be needed. Caregiving and community contributions should receive recognition without making subsistence conditional on proving moral worth.
For relationships, the opportunity is more time together; the danger is insecure income combined with isolation. Family policy should support shared caregiving and predictable schedules, rather than allowing paid-work reductions to become another unequal domestic burden.
AI companionship may help some people, but should not become the default substitute for human connection. A four-week study of 981 participants found that heavier voluntary chatbot use was associated with worse psychosocial outcomes; the study did not randomly assign heavy use and therefore does not establish that heavy use caused those outcomes.
My policy recommendation is clear disclosure, strong privacy protection, safeguards for minors, and restrictions on designs that deliberately cultivate dependency or discourage human relationships.
Human worth cannot sensibly depend on outperforming a machine. That is a principle society must choose, not an empirical prediction that technology will deliver automatically.
PhaseMain actionsEvidence needed before expansionFirst 100 daysEstablish a labor-transition observatory; stress-test benefit delivery; track hiring, hours, earnings, and contractor income; identify exposed regions and entry pathwaysBaseline data, delivery capacity, coverage gapsFirst yearPilot paid transitions, wage insurance, negotiated work-sharing, and employer-linked apprenticeships; strengthen minimum-income accessReemployment quality, disposable income, service quality, participation, costYears 2–3Scale successful programs; reform revenue collection; build ownership mechanisms; expand housing, energy, and care capacitySustained gains rather than initial placement counts; credible recurring financingIf displacement persistsExpand permanent income floors and dividends, consider further reductions in working time, reduce inappropriate employment conditions on supportPersistent labor-demand weakness and distributional deterioration, not headline speculation
Use precommitted triggers
Governments should announce in advance which conditions activate stronger support. Illustrative triggers—not validated universal thresholds—could include:
A sustained two-percentage-point decline in prime-age employment, or a 15% deterioration in entry-level hiring relative to an appropriate comparison, could trigger additional transition funding and extended income support.
A persistent decline in labor’s income share while productivity rises could trigger a review of taxation, ownership, and gain-sharing arrangements.
Rapid increases in essential living costs should trigger attention to supply constraints rather than automatically assuming a larger nominal transfer solves the problem.
Thresholds need local calibration, regional breakdowns, and protection against statistical noise.
Evaluate the policies, not merely the participants
Where ethically feasible, use randomized access to additional services or phased rollouts across comparable firms and regions. Basic safety-net rights should not depend on assignment to an experiment.
Measure outcomes after 12 and 24 months: earnings, disposable income, housing stability, job quality, health, social participation, service quality, and public cost. Include people who never obtained a first job, not only those formally laid off.
For work-sharing, check whether fewer hours become more intense hours. For training, check whether earnings gains persist. For cash, examine purchasing power and interactions with housing and care costs. For ownership funds, examine governance and who actually receives the returns.
Stop or redesign programs that fail. Do not label every unsuccessful course “worker resistance,” or every missed fiscal target a temporary problem.
Recommendation
Adopt an adaptive shared-prosperity strategy: a dependable income floor and essential services immediately; paid transitions and work-sharing while useful employment remains plentiful enough; and progressively broader capital ownership and social dividends if labor becomes a less reliable source of income.
This is the strongest path because it does not depend on one prediction being correct.
With modest disruption, it improves mobility and security without requiring a wholesale replacement of employment. With rapid productivity growth, it distributes benefits through income, services, ownership, and time. With weak growth and displacement, it reveals the need for broader financing rather than pretending an AI windfall will pay every bill.
Its success should be judged by real living standards, economic security, meaningful choice, social connection, and the distribution of power—not simply by GDP growth or the number of hours humans can still be paid to perform.
The objective is not to preserve every existing job. It is to ensure that losing the economic necessity of some human labor does not mean losing the economic rights, social standing, or purpose of the people who performed it.
Comments (0)
Please sign in to leave a comment.
No comments yet. Be the first to share your thoughts!
