David Aterido, secretary of labor for a state of nine million people, opened a briefing he had requested but did not want to read. Over the next decade, the analysis estimated, AI would significantly change the tasks in roughly 40 percent of jobs across the state, with three occupational clusters facing the steepest disruption: customer service, data entry and processing, and entry-level paralegal and clerical work. Those clusters employed about 280,000 of his residents. The same analysis projected new demand in roles the state's training programs barely touched. David's predecessors had treated workforce policy as a slow, generational lever. He realized he had, at most, a few budget cycles before the disruption arrived faster than his institutions could respond. The question was not whether to act. It was how to move education, training, and labor policy at the speed AI was moving the economy.

National and state leaders own a problem most agency heads can avoid: AI does not just change how their own agency works, it reshapes the entire labor market they are responsible for, and it reshapes their own workforce at the same time. This lesson treats workforce transformation as a policy design challenge, carries David's 280,000 at-risk workers as its spine, and gives leaders a structure for aligning education, retraining, labor policy and their own employment practices with a moving target.

The Real Shape of the Disruption

The public debate fixates on "AI takes jobs." David's analysts gave him a more useful frame: AI changes tasks faster than it eliminates whole jobs. A paralegal does not vanish; the document-review part of the job is automated, and the role shifts toward judgment, client interaction, and supervising AI output. This distinction is the whole policy lever. If jobs simply disappeared, the only response would be income support. Because tasks shift, the response can be transition: helping workers move into the parts of their work, or adjacent work, that AI does not do.

The national evidence points the same way, though the numbers measure different things and should not be blurred together. Analyses cited in the source material, drawing on Bureau of Labor Statistics work and the OECD 2023 Employment Outlook, estimate that roughly 27 percent of jobs in advanced economies are at high exposure to generative AI, with additional tasks subject to augmentation rather than replacement. That is an exposure estimate across advanced economies. David's 40 percent is a task-change estimate for one state over one decade. Both are projections with error bars, and neither is a headcount of jobs that will disappear.

What the estimates agree on is that the distribution is uneven. Clerical, customer service and paralegal roles see disproportionate task exposure. That matters beyond economics, because those roles are disproportionately held by women and workers of color. Three consequences follow for policy. The disruption is concentrated in specific clusters and regions, so policy should be targeted rather than blanket. It is fast, outpacing the multi-year cadence of curriculum reform. And it is continuous; this is not one wave but a standing condition, which means one-time retraining programs will fail by design rather than by mismanagement.

Government Is Not a Spectator: It Is a Large Employer Deploying AI

A workforce policy leader who talks only about the private labor market will be caught out by their own payroll. Agencies across the federal government, including the General Services Administration, the Department of Homeland Security, the Department of Veterans Affairs, the Internal Revenue Service, the Social Security Administration and the Department of Defense, are deploying generative AI assistants for drafting, customer contact and case triage. The Veterans Affairs department has piloted AI-assisted claim processing against an annual disability claim volume the source puts at more than 1.3 million. The IRS has piloted generative AI for taxpayer correspondence.

Each of those deployments reshapes jobs, and each therefore requires training, change management and consultation with bargaining units. The obligations here are legal, not cultural. The Federal Service Labor-Management Relations Statute requires that covered changes be negotiated with certified bargaining units, and the source names the American Federation of Government Employees and the National Treasury Employees Union as the relevant units in the deployments it describes. OMB Memorandum M-24-10 requires rights-impacting AI to include worker consultation and human oversight where workforce decisions are affected.

Treat those obligations as strictly as the source states them and then verify the specifics with your labor relations office and counsel, because what is negotiable, what is merely consultable, and what timing applies are questions with real legal answers that vary by change and by unit. What is not variable is the sequencing lesson. Consultation that begins after the vendor is selected and the workflow is redesigned is not consultation; it is notification with extra meetings. The deployments the source treats as constructive are the ones where the bargaining unit was in the room while the design was still capable of changing.

The Four Levers a Leader Actually Controls

David mapped his external policy options to four levers, each with a different timescale and owner.

• Education alignment. What the school and community-college system teaches. Slowest to change, highest long-term leverage. The goal is durable skills, judgment, communication, and the ability to work alongside AI, over narrow technical content that ages fast.
• Retraining and reskilling. Programs for adults already in the workforce. Faster, and where David's 280,000 at-risk workers live. The goal is short, stackable credentials tied to real local demand.
• Labor market policy. Unemployment systems, wage insurance, portable benefits, and transition support for workers between roles.
• Demand signaling. Using the state's own hiring, procurement, and incentives to grow the jobs the transition needs, so retraining points somewhere real.

The Instruments Already on the Shelf

Leaders reach for new programs when the harder and cheaper move is to point existing instruments at the new problem. On the training side, the Workforce Innovation and Opportunity Act provides adult, dislocated worker and youth funding through local workforce development boards under Title I, with Title II covering adult education, and it carries its own performance metrics. Trade Adjustment Assistance supports trade-affected workers, and the source notes that advocates have pushed for an AI analog without asserting that one exists. Registered Apprenticeship under the Department of Labor has expanded into AI roles through the Apprenticeship USA initiative.

Other instruments reach populations that general programs miss: YouthBuild, Job Corps, the Senior Community Service Employment Program, and veterans' Transition Assistance Programs. Public delivery runs through the American Job Center one-stop system alongside unemployment insurance modernization. On the education side, Perkins V anchors career and technical education, Pell Grants now include short-term programs under limited pilots, Department of Education competency frameworks and the National Educational Technology Plan shape state plans, and NSF-funded AI institutes seed the research pipeline. Community colleges remain the workhorses of mid-career retraining.

Education alignment is the slowest instrument and the one most often promised and least often delivered, because it is a decade-long project measured against annual budgets. The pieces are already in motion: K-12 computer science standards, advanced placement computer science, NSF-funded AI institutes, competency-based assessment, and postsecondary work on recognising learning wherever it happens. Community colleges do most of the mid-career work, and the source notes AI-specific programs documented at hundreds of institutions by their national association. Nothing here moves on a disruption timescale, which is exactly why it has to be started before the disruption is visible in the enrollment data.

On the response side, the source groups policy into four clusters, and the grouping is a useful test of whether your own plan is balanced. Modernizing the safety net, primarily through unemployment insurance reform. Expanding and targeting training investments. Requiring transparency and worker voice in AI deployment. And building institutional capacity for workforce impact assessment, so that the effect of a deployment is something the government can assess rather than something it learns about from the people it happened to. Most jurisdictions are strong in the second cluster and thin in the fourth, which is why so many programs run without anyone able to say whether they worked.

The division of federal labor is worth knowing before you ask the wrong department for the wrong thing. Labor leads training and unemployment insurance. Commerce leads standards and measurement. Education leads postsecondary alignment. OMB leads federal procurement and AI governance. NIST leads risk management frameworks. GAO provides accountability oversight. State labor departments, community college systems and local workforce boards execute. Almost every real coordination failure in this space is a seam between two of those owners, not a gap inside any one of them, which is why naming a single cross-agency owner matters more than any individual program design.

Designing Retraining That Does Not Fail

David knew the graveyard of government retraining: programs that trained people for jobs that did not exist, or for skills already obsolete by graduation. He set three design rules drawn from what survives. First, train to verified local demand, not forecasts. Rather than guessing future jobs, the program partnered with employers who committed to interview or hire graduates. A credential with no employer on the other end is a cruelty, not a service.

Second, short and stackable. Twelve-week credentials that build toward larger qualifications, so a displaced data-entry worker can earn an early win and keep climbing, rather than facing a two-year program they cannot afford to enter. Third, continuous, not one-shot. Because the disruption is ongoing, David funded a standing reskilling capacity rather than a single appropriation, treating it like infrastructure rather than a project. State models pointing the same direction include California's High Road Training Partnerships, New York's Office of Strategic Workforce Development, and Colorado's Opportunity Now initiative.

One honest caveat about the first rule. An employer commitment to interview graduates is a far weaker instrument than it sounds, and even a commitment to hire is a commitment under this year's conditions. Verified demand reduces the risk of training people for nothing. It does not transfer the labor market's risk onto the employer, and a program that reports placement rates without reporting retention and wages is measuring the easy half. Design the follow-up before you design the curriculum.

Seeing the Disruption Before It Lands

None of the levers work without intelligence about where the disruption is actually falling. National baselines come from Bureau of Labor Statistics occupational employment and wage statistics, the Job Openings and Labor Turnover Survey, and the American Community Survey. State labor market information offices add local detail. The source notes federal calls for more granular measurement, and points to the GAO AI Accountability Framework's treatment of performance monitoring as a governance obligation rather than an optional analytic nicety. Workforce boards and state labor departments need dashboards that combine national survey data with faster local signals.

Build that capacity with clear eyes about what it can do. Survey instruments were designed to describe an economy that changes slowly, and they report on a lag. A dashboard shows you the disruption you decided to instrument, in the categories your data already had, and the clusters that surprise you will be the ones nobody built a field for. Monitoring aligned to the measurement function of a risk framework is a discipline for noticing sooner. It is not an early warning system, and treating a green dashboard as evidence that the transition is going well is how a leader misses the first two years of a shift.

Equity Is Both the Whole Game and a Legal Obligation

David's analysts flagged a danger that aggregate optimism hides. The workers most exposed to AI disruption, in clerical, customer-service and data-processing roles, are disproportionately lower-wage, and often women and workers of color. If retraining programs are accessed mainly by those with time, money and digital access, AI transformation widens inequality even as the state's average outcomes improve. David therefore required every program to report participation and completion disaggregated by income, geography, language and disability, and to meet Section 508 accessibility standards so the tools delivering the training did not become their own barrier.

The policy anchor the source keeps returning to is the Blueprint for an AI Bill of Rights, published by the White House Office of Science and Technology Policy as a non-binding framework, and specifically its principle on algorithmic discrimination protections. Treat it as what it is: a statement of expectations that shapes how agencies and oversight bodies read the binding obligations, not a source of enforceable rights on its own. Its practical value in workforce settings is that it names the failure before the litigation does, which gives a programme manager language for a concern that would otherwise have no vocabulary until a complaint arrives.

Where AI is used on workers rather than for them, the obligations harden from good practice into law. The source is direct that Equal Employment Opportunity Commission guidance on AI hiring tools, disability accommodations and selection procedures carries Title VII, Americans with Disabilities Act and Age Discrimination in Employment Act obligations, that the Department of Labor's Office of Federal Contract Compliance Programs enforces affirmative action obligations for federal contractors, and that disparate impact testing, reasonable accommodation protocols and transparency about AI use in hiring and promotion are legal necessities rather than options. Carry that framing strictly, and confirm the current guidance and its application with counsel before you rely on it.

State law is moving too, unevenly. The source names the Illinois Artificial Intelligence Video Interview Act, Colorado employment protections addressing AI, and California proposals under AB 2930, keeping the last of those explicitly at the proposal stage. If you operate in multiple jurisdictions, the practical consequence is that a single national hiring tool may face different disclosure and testing obligations in different states, and the compliance question is not whether the vendor is compliant but whether your use of it is, in each place you use it.

Two cases show what failure looks like from the citizen's side. A federal identity-verification rollout built on biometric enrollment, later reversed, excluded veterans, older Americans and citizens without smartphones because accessibility was not tested before deployment. And in decisions that affect livelihoods, the source points to the COMPAS case and the Houston school district teacher evaluation litigation as illustrations of due process and discrimination risk. Disaggregated reporting is how you find these before a court does. It is worth being blunt that measuring a gap does not close it; a report that is produced, filed and never acted on has changed nothing except the paper trail.

What Constructive and Destructive Deployment Look Like

The source offers paired examples, and the pairing is the lesson. On the destructive side, Michigan's MIDAS unemployment fraud system automated adjudication at scale and produced false accusations that took years to unwind. The Dutch childcare benefits scandal did analogous damage in another country, where automation displaced judgment without adequate workforce or citizen protection. In both, the technology was the visible failure and the absence of human oversight, appeal capacity and trained adjudicators was the actual one.

On the constructive side, the same source describes Michigan's post-MIDAS reforms, which combined independent review, human oversight and expanded adjudicator training; Allegheny County's reskilling of child welfare workers around the screening tool they were being asked to use; and federal deployments where the AI assistant rollout and the customer service pilot each proceeded with the relevant bargaining unit consulted. The common factor is not that the constructive cases used better models. It is that governance, bargaining and training were designed together, so the people whose jobs changed were equipped and represented rather than merely informed.

A Workforce Transformation Readiness Rubric

David built a rubric to assess, and report on, how ready his state's workforce system actually was. Score each dimension from 1 (reactive) to 4 (anticipatory). The pattern reveals where to invest next, and gives the leader something to track over time.

1. Labor-market intelligence. 1: We react to layoffs after they happen. 4: We forecast task-level disruption by cluster and region and update it continuously.
2. Education alignment. 1: Curricula change on a multi-year cycle disconnected from AI shifts. 4: Durable-skills and AI-collaboration content is embedded and refreshed regularly.
3. Retraining design. 1: Long, generic programs with no employer link. 4: Short, stackable credentials tied to committed employer demand, with retention and wage follow-up.
4. Speed. 1: New programs take years to launch. 4: We can stand up a targeted reskilling track within a single budget cycle.
5. Equity of access. 1: Programs reach the already-advantaged. 4: We measure and act on gaps by income, geography, language and disability, meeting accessibility standards.
6. Safety net. 1: Support ends at unemployment benefits. 4: Wage insurance, portable benefits and transition support bridge workers between roles.
7. Worker voice. 1: Bargaining units learn about AI deployments when they launch. 4: Consultation happens while the design can still change, and the obligation is verified with counsel.
8. Funding model. 1: One-time appropriations. 4: Standing, infrastructure-grade reskilling capacity.
9. Governance. 1: No one owns workforce-AI strategy. 4: A named leader and cross-agency body own it, with public reporting.

International Benchmarks and Their Limits

Several countries have built the standing capacity this lesson argues for, and their designs are worth studying. Singapore's SkillsFuture provides credits inside a lifelong learning architecture. Germany's dual education system links employers and vocational education structurally. The United Kingdom's Lifelong Learning Entitlement restructures postsecondary funding toward modular study. Canada's Future Skills Centre funds and evaluates workforce experimentation. Australia has expanded apprenticeships, and South Korea's Digital New Deal linked digital investment to workforce development.

The source is emphatic about how to use these, and the warning applies to every borrowed policy in this course. Each reflects a distinct political economy and cannot be copied without adaptation. Analyse them for transferable principles rather than copyable blueprints, with explicit attention to institutional context and political feasibility. A credit system works in a country with the administrative infrastructure to run it and the employer relationships to make the credits mean something; imported without either, it becomes a voucher program with low take-up and a disappointed evaluation three years later.

The Leadership Move

David's core decision was to treat workforce transformation as continuous infrastructure rather than an episodic crisis response. He named a cross-agency owner spanning labor, education and economic development, stood up standing labor-market intelligence so the state could see disruption coming, and tied retraining to verified demand. The technology will keep moving; the institutions that respond well are the ones built to keep moving with it. The aim is not to predict exactly which jobs AI changes, an impossible task, but to build a workforce system fast and fair enough to help people transition no matter which way the disruption breaks.

For the agency's own deployments, the source sets out the shape of a two-year plan, and it is a useful spine: baseline assessment, identification of the impacted population, bargaining unit consultation, training design, vendor selection aligned with federal cloud authorization and governance requirements, communications, metrics, and governance, with explicit go or no-go gates, equity monitoring, and escalation paths for incidents. Note where consultation sits in that sequence. It is early, before training design and vendor selection, which is the only position from which it can change anything.

Anti-Patterns

• Treating retraining as a guarantee of transition. Funded training reduces the risk that a displaced worker has nowhere to go. It does not place anyone. A program that reports enrollments and completions but not placement, retention and wages is reporting effort as if it were outcome, and will look successful for exactly as long as nobody follows up.
• Mistaking an employer commitment for a job. A commitment to interview is not a commitment to hire, and a commitment to hire is made under this year's conditions. Build the follow-up and the contingency before you build the curriculum.
• Consulting the bargaining unit after the design is fixed. Notification presented as consultation satisfies nobody, changes nothing, and creates a legal exposure on top of the workforce one. If the answer to "what could this conversation change" is nothing, the conversation is theatre.
• Letting disaggregated reporting substitute for action. Measuring a gap by income, geography, language or disability tells you it exists. Closing it requires someone with a budget who is accountable for the number moving. A completed equity report has never once enrolled anybody.
• One-time appropriations for a continuous condition. A single reskilling appropriation buys one cohort against a standing disruption, then leaves a program with staff, no funding and a waitlist.
• Copying another jurisdiction's program wholesale. Every national and state model rests on institutions, employer relationships and political economy that do not travel. Take the principle; rebuild the mechanism locally.
• Trusting the dashboard to warn you. Labor market instruments report on a lag, in categories designed for a slower economy, about the things you chose to instrument. Monitoring helps you notice sooner. It does not tell you what you failed to measure.
• Confusing exposure estimates with job-loss counts. Task-exposure projections describe how much of a role could change, across different populations and time horizons, with wide uncertainty. Quoting one as a headcount of eliminated jobs will misdirect a budget and, once corrected in public, cost you the credibility you needed to act.

Practice Prompts

• Take the three occupational clusters most exposed in your jurisdiction and, for each, separate the tasks AI is likely to change from the whole roles it might eliminate. Which parts of the work does the automation not touch, and what would it take to move a worker toward those parts?
• Inventory the workforce instruments already available to you, federal and state, and map which population each actually reaches. Where are the gaps between the instrument's eligibility rules and the people your exposure analysis identified?
• Draft the consultation plan for one AI deployment inside your own organization. Name the bargaining unit or employee representatives, the point in the timeline where consultation happens, and the specific design decisions still open at that point.
• Design a single reskilling track end to end for one exposed cluster: employer partners, credential structure, duration, funding source, accessibility provisions, and the follow-up measurement of placement, retention and wages.
• Score your jurisdiction against the readiness rubric and identify the two lowest dimensions. For each, name the owner, the first move, and what it would cost to reach the next level.
• Pick one international model and write the honest version of why it would not transfer to your context unchanged. Then extract the one principle underneath it that would.

Reflection

Think about the last time your organization automated part of someone's job. Who found out, and when? Was the person whose work changed consulted while the design could still move, or informed once the decision was made? What training did they receive, and who paid for it? Now ask the uncomfortable follow-up: if you had to defend that sequence to the worker, to their representative, and to an oversight body, which part would you defend confidently and which part would you find yourself explaining?

Then consider the wider obligation. Your jurisdiction's most exposed workers are, on the evidence in this lesson, disproportionately lower-wage and disproportionately women and workers of color. Whatever you build over the next few budget cycles will either reach them or reach the people who already know how to navigate public programs. That outcome is not decided by intent. It is decided by eligibility rules, application design, program hours, language access and accessibility, all of which are choices someone in your organization is making right now, mostly without thinking of them as equity decisions.

Glossary

• Task exposure. The share of the activities within an occupation that a technology could plausibly perform or substantially change, as distinct from the share of jobs eliminated.
• Augmentation. Technology changing how a task is performed while the worker remains central to it, contrasted with substitution.
• Stackable credential. A short qualification designed to combine with others toward a larger certificate or degree, so learners gain value at each step.
• Dislocated worker. A worker who has lost employment through no fault of their own and is eligible for specific transition programs under federal workforce law.
• Registered Apprenticeship. An earn-while-you-learn training model registered with the Department of Labor, combining paid work with structured instruction.
• Bargaining unit. A group of employees represented by a certified labor organization, whose representation triggers specific obligations when working conditions change.
• Disparate impact. A facially neutral practice that produces significantly different outcomes across protected groups, analysed under specific legal frameworks that are not identical to each other.
• Wage insurance. Transitional support that partially replaces lost earnings when a worker moves into a lower-paying role rather than into unemployment.
• Portable benefits. Benefits attached to the worker rather than to a single employer, so they survive a job change.
• Labor market information. The survey and administrative data infrastructure describing employment, wages, openings and turnover, produced nationally and supplemented by state offices.

Related Lessons

• Workforce Transformation at Scale takes the same problem from inside a single large organization rather than a whole labor market.
• AI Talent Pipeline: Education System Alignment develops the education lever in depth.
• AI and Equity: Reaching All Communities extends the access and disaggregation discipline across service delivery generally.
• Workforce Planning for AI covers the agency-level staffing and skills mix this lesson touches only from the policy side.
• AI in Social Services examines the benefit systems that absorb the people a transition fails.
• Change Management for AI Adoption supplies the practice behind the consultation and training sequencing described here.
• AI and the Social Contract asks what government owes the people its own automation decisions displace.
• National AI Competitiveness places workforce capacity inside the wider strategic picture.

Closing

Workforce transformation is the rare policy problem where the government is simultaneously the regulator, the funder, the trainer and a large employer doing the automating. Those four roles pull in different directions, and a leader who occupies only one of them will produce policy that the other three quietly undermine. The coherence has to be built deliberately, through a named owner, shared intelligence and a plan that treats the state's own deployments by the same standard it sets for everyone else.

David's insight was not about technology. It was about tempo. His institutions were built to move on generational timescales against a disruption moving on budget-cycle timescales, and no amount of forecasting closes that gap. What closes it is standing capacity: intelligence that runs continuously, funding that does not expire, credentials that can be stood up in months, and consultation that happens early enough to matter. Build those, and you do not need to predict which jobs change. You need only be ready when they do.

Key Takeaways

• AI changes tasks faster than it deletes jobs. That distinction is the central policy lever: it lets leaders build transition bridges instead of only catching people who fall.
• Do not blur exposure estimates with job losses. Different studies measure different populations over different horizons, and quoting a projection as a headcount will misdirect a budget and cost you credibility.
• The disruption is targeted, fast and continuous. Policy should be cluster-specific, faster than curriculum cycles, and standing rather than one-shot.
• Government is also the employer. Agency AI deployments reshape jobs and carry legal consultation and oversight obligations, and consultation only counts when the design can still change.
• Use the instruments you already have. Federal and state workforce, apprenticeship and education programs exist; the usual failure is a seam between owners, not a missing program.
• Train to verified demand, then follow up. Employer-linked, short and stackable beats long and generic, but an interview commitment is not a job and placement is not retention.
• Fund reskilling like infrastructure. Continuous disruption demands standing capacity, not episodic appropriations.
• Make equity measurable and then act on it. The most AI-exposed workers are often the most vulnerable, disaggregated reporting is how you find the gap, and only a budget and an owner close it.
• Borrow principles, not blueprints. International models rest on institutions that do not travel; rebuild the mechanism locally.
• Name an owner. A cross-agency leader with public reporting turns workforce-AI strategy from everyone's concern into someone's accountability.

Frequently Asked Questions

How much of the workforce is actually at risk? No honest answer is a single number. The source material cites an estimate that roughly 27 percent of jobs in advanced economies are at high exposure to generative AI, with further tasks subject to augmentation, while the state-level analysis in this lesson estimates significant task change in roughly 40 percent of jobs over a decade. Those measure different things over different populations. Use them to identify where to look, not to forecast a headcount, and always carry the attribution and the hedge with the figure.

Do we have to bargain with a union before deploying an AI tool? The source states that the Federal Service Labor-Management Relations Statute requires covered changes to be negotiated with certified bargaining units. Whether a specific deployment is a covered change, and what obligations attach, is a legal question for your labor relations office and counsel, and the answer varies by change and by unit. Do not resolve it by reading a course. Do resolve the sequencing yourself: bring representatives in while the design is still open, which is good practice regardless of what the legal minimum turns out to be.

Is retraining actually effective, or is it a political reflex? The evidence in this lesson supports a narrower claim than "retraining works." Programs tied to verified local demand, built as short stackable credentials, and funded continuously survive better than long generic programs built on forecasts. That is a comparison between designs, not a promise of outcome. Any program that cannot report placement, retention and wages after the fact has no evidence about its own effectiveness, only about its throughput.

What is the single most common failure mode? Treating the transition as an event. One-time appropriations, one-time analyses, one-time consultations and one-time equity reports all assume a disruption that arrives, is handled, and ends. The disruption in this lesson is a standing condition, and every institution built for an event will be out of money, out of data or out of legitimacy at roughly the same moment the second wave arrives.

Where should a leader start if they have one budget cycle? Two moves, in order. Name a cross-agency owner with authority spanning labor, education and economic development, because almost every failure here lives in a seam. Then stand up continuous labor-market intelligence, however modest, because every other decision depends on knowing where the disruption is falling. Retraining tracks, safety net changes and education alignment all take longer than a cycle; ownership and visibility do not.

How do we avoid the transformation widening inequality? By deciding that reach is an outcome you are accountable for, not a hope. Report participation and completion disaggregated by income, geography, language and disability, meet accessibility standards in the tools that deliver the training, and give someone a budget and a target for closing whatever gaps the reporting exposes. Programs left to their own dynamics reach the people who already know how to find public programs, which is precisely not the population this policy exists to serve.