---
title: "Care Workforce Surge — Projection Validation Record"
status: "Evidence review"
version: "0.1"
date: "2026-08-19"
applicability: "Research record. It validates or qualifies a strategic projection; it creates no capability, service, institutional authority, or clinical claim."
---

# Care Workforce Surge — Projection Validation Record

## 1. The projection under review

Stated in full, as received:

> Healthcare systems will face an influx of workers transitioning into caregiving roles. A Nurse AI OS will be critical infrastructure to upskill, manage, and coordinate this rapidly growing workforce. It must be designed not just for veteran nurses, but as an intuitive co-pilot to help train and scale newer human-to-human care workers seamlessly.

It contains three separable claims. They do not stand or fall together, and this record grades them separately.

| # | Claim | Verdict | Confidence |
|---|---|---|---|
| **C1** | Healthcare will face an influx of workers transitioning into caregiving roles | **Supported, with the mechanism corrected** — the pull is demographic demand, not AI displacement; and the influx arrives as churn, not as accumulated tenure | High for the flow; moderate for its composition |
| **C2** | A Nurse AI OS will be critical infrastructure to upskill, manage, and coordinate that workforce | **Need supported; status aspirational; "manage" must be narrowed** — the coordination gap is real, but no such system holds critical-infrastructure status today, and workforce *management* runs into employment-decision and title-protection law | Moderate for the need; the status claim is unevidenced and should not be made |
| **C3** | It must be designed as an intuitive co-pilot for newer care workers, not only veterans | **Supported, and it is the strongest of the three — but the naive version of it is contradicted by evidence** | High, with a hard design constraint attached |

The rest of this record shows the work.

---

## 2. C1 — Is the influx real?

### 2.1 What the demand side says

The pull is not speculative. It is the single largest occupational movement in the U.S. projections.

- Home health and personal care aides are already **the largest occupation in the economy, about 4.3 million jobs in 2024**, and are projected to **add 739,800 jobs (17.0 percent) between 2024 and 2034 — the most of any of the 832 detailed occupations, roughly 1 of every 7 jobs added over the decade.** The second-largest gainer, software developers, adds 267,700 ([BLS, *Monthly Labor Review*, 2024–34 projections overview](https://www.bls.gov/opub/mlr/2026/article/industry-and-occupational-employment-projections-overview.htm)).
- Total employment over that decade grows 3.1 percent, from 170.0 million to 175.2 million — **5.2 million new jobs economy-wide, of which health care and social assistance supplies roughly 2.0 million** (same source).
- The Occupational Outlook Handbook projects about **765,800 openings per year** for home health and personal care aides ([BLS OOH](https://www.bls.gov/ooh/healthcare/home-health-aides-and-personal-care-aides.htm)).
- Counting exits and transfers as well as growth, PHI puts the total at **an estimated 9.7 million direct-care job openings between 2024 and 2034** ([PHI, *Direct Care Workers in the United States: Key Facts 2025*](https://www.phinational.org/resource/direct-care-workers-in-the-united-states-key-facts-2025/)).

The near-term data is even starker than the projection. **Health care and social assistance employment rose 2.9 percent — 680,500 jobs — from March 2025 to March 2026, in a year when total nonfarm payroll employment "changed little on net"** ([BLS, *The Economics Daily*](https://www.bls.gov/opub/ted/2026/health-care-and-social-assistance-employment-increased-by-2-9-percent-or-680500-from-march-2025-to-march-2026.htm)). One sector is carrying national job growth while the rest of the economy is flat. Secondary analyses of the same period put health care at more than 100 percent of net job creation in some months — arithmetically possible only because other sectors shrank.

Globally the same shape appears: the nursing workforce grew from 27.9 million (2018) to 29.8 million (2023) while the shortage fell from 6.2 million to 5.8 million, projected to reach 4.1 million by 2030 — with roughly 70 percent of the remaining shortage concentrated in the WHO African and Eastern Mediterranean Regions, and 78 percent of the world's nurses working in countries holding 49 percent of the world's population ([WHO, *State of the World's Nursing 2025*](https://www.who.int/publications/b/78362)).

### 2.2 What the supply side says

The projection implies workers *transitioning in* from elsewhere. There is real pressure producing candidates:

- Stanford Digital Economy Lab's "canaries" work, using ADP administrative data, finds employment for **workers aged 22–25 in highly AI-exposed occupations running roughly 19 percent below where it would be had it kept pace with less-exposed occupations** as of mid-2026 — with the effect concentrated where AI automates rather than augments, and **no evidence of widespread economy-wide displacement** ([Stanford DEL](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/)).

That is a real supply-side push. But it is honest to state what the evidence does **not** show: there is, at this writing, **no published measurement of a displaced-worker-to-caregiving migration**. The two facts — young workers squeezed out of AI-exposed white-collar entry points, and caregiving absorbing national job growth — are adjacent, not causally linked in any study reviewed here. Anyone who claims the link is measured is overstating it. What is defensible is that the destination has capacity for millions of entrants and the alternatives for entry-level workers are narrowing.

### 2.3 The correction that matters

The word "influx" implies a stock filling up. The evidence describes a **flow that mostly drains back out**.

- Median hourly wage for direct care workers was **$17.36 in 2024**, with **median annual earnings just under $26,000** — low wages compounded by part-time hours (PHI Key Facts 2025).
- **Home care turnover was near 75 percent in 2024**; median annual turnover for nursing assistants in nursing homes was **near 100 percent** in the 2017–2018 measurement (PHI).
- **Immigrants are 28 percent of the direct-care long-term-care workforce and about 30 percent of aides — over 820,000 people — and 32 to 40 percent of home care workers** ([KFF](https://www.kff.org/medicaid/what-role-do-immigrants-play-in-the-direct-long-term-care-workforce/)). Modeled deportation scenarios put potential direct-care job losses at **394,000** (274,000 immigrant workers plus 120,000 complementary roles) ([PHI immigration brief](https://www.phinational.org/immigration-and-the-direct-care-workforce/)). That is a scenario, not a forecast — but it sizes the exposure.

Meanwhile the licensed lane cannot expand to absorb the flow, because it is capacity-capped rather than demand-capped:

- U.S. nursing schools turned away **93,176 qualified applications in 2025–26**, the highest on record.
- National nurse faculty vacancy rate **8.8 percent**, more than 1,600 unfilled positions; **more than 60 percent of programs cite lack of clinical placement sites** as a barrier to growth (AACN survey data, as reported in [nursing workforce coverage](https://www.medscape.com/viewarticle/nursing-faculty-shortage-limiting-rn-pipeline-2026a1000nmb)).

And the experienced nurses who would supervise the entrants are themselves leaving:

- **39.9 percent of RNs and 41.3 percent of LPN/VNs report intent to leave the workforce or retire within five years** (18.6 percent retiring, 22.7 percent leaving for other reasons); **median age 50**; about **41.5 percent name stress and burnout** as the root cause; and roughly **188,962 RNs under 40** report the same intent ([NCSBN, *2024 National Nursing Workforce Survey*, published 2025](https://www.journalofnursingregulation.com/article/S2155-8256(25)00047-X/fulltext)).

### 2.4 Restated projection

> Healthcare is not about to receive a wave of new caregivers who stay. It is settling into a **permanent high-volume, low-tenure, mixed-provenance workforce** — millions of entries and near-equal exits per decade, an unusually large share of hands-on care delivered by people with under twelve months in role, and a shrinking, aging pool of experienced nurses expected to orient, delegate to, verify, and rescue all of them.
>
> **The binding constraint is not labor supply. It is supervisory capacity per experienced nurse.**

That restatement is the single most consequential output of this review, because it changes what the software is for. A system built to *add capacity* competes with the wage problem and loses. A system built to *multiply supervision* addresses the actual bottleneck and does not pretend to fix pay.

---

## 3. C2 — Is a Nurse AI OS critical infrastructure for this?

**The coordination need is real.** Entrants arrive through a training system with a low and uneven floor: the federal minimum for nurse aide training is **75 clock hours including at least 16 hours of supervised practical training** ([42 CFR 483.152](https://www.ecfr.gov/current/title-42/chapter-IV/subchapter-G/part-483/subpart-D/section-483.152)); 30 states and D.C. exceed it, up to 180 hours, 13 states and D.C. require 120+ — and roughly **20 states have not changed their requirement in nearly 30 years** ([PHI training requirements tracker](https://www.phinational.org/advocacy/nurse-aide-training-requirements/)). A worker's preparation therefore depends heavily on which state they entered in and which employer trained them. Something has to carry orientation, scope, escalation, and local practice across that variance. Today that something is an experienced nurse's attention, and there is not enough of it.

**The status claim should not be made.** Nothing in this repository, and no system known to this review, is critical infrastructure for the care workforce. Publication does not confer that status; adoption, dependency, and evidence would, and none exist yet. Under this project's own doctrine — *no fabricated evidence, review, approval, credential, conformance, or institutional status* — "will be critical infrastructure" is a goal to be earned, not a description to be published.

**"Manage" must be narrowed, for legal reasons and not merely stylistic ones.**

- **Title protection.** A growing set of states bar nonhuman entities from holding licensed clinician status or protected titles. Delaware HB 191 prohibits nonhuman entities, including AI agents, from being licensed or certified as nurses, APRNs, practical nurses, physicians, or physician assistants, and bars use of the protected titles ([Holland & Knight, 2026 state AI health legislation review](https://www.hklaw.com/en/insights/publications/2026/05/states-continue-efforts-to-regulate-ai-in-healthcare)). Oregon HB 2748 (2025, effective January 1, 2026) and California AB 489 (2025, effective January 1, 2026) reach the same result by different routes; Washington HB 2155 was signed March 9, 2026, effective June 11, 2026 ([Nurse.org statehouse AI tracker](https://nurse.org/news/nursing-ai-legislation-tracker/)). A co-pilot for new care workers must never present, be described, or be perceived as a nurse.
- **Consequential decisions.** The same legislative wave repeatedly requires that adverse or clinical determinations be made by licensed humans rather than dictated by a model (e.g., Indiana HB 1271, Utah SB 319, Georgia SB 544, Washington SB 5395, per the same review). Workforce "management" — competency determination, discipline, scheduling penalties, promotion, hiring — is the employment analogue of that same category. A system that makes those calls inherits that regulatory surface and the liability with it.

**Narrowed claim that survives review:** a Nurse AI OS can be *supporting* infrastructure for orientation, scope-awareness, escalation routing, and documented supervision — proposing, never deciding, and never wearing a license.

---

## 4. C3 — Should it be built novice-first?

Yes, and the evidence for it is the strongest in this record — which is exactly why the naive implementation is dangerous.

### 4.1 The case for novice-first

The best-identified field evidence on generative AI assistance finds the gains land where this projection says they are needed. In a staggered rollout across **5,172 customer-support agents**, access to an AI assistant raised issues resolved per hour by **15 percent on average**, with **less-experienced and lower-skilled workers improving both speed and quality**, while **the most experienced workers saw small speed gains and small quality declines**; the study also reports improved customer sentiment, improved employee retention, and signs of worker learning ([Brynjolfsson, Li & Raymond, *Generative AI at Work*, QJE 140(2), May 2025](https://academic.oup.com/qje/article/140/2/889/7990658)).

Two things in that result matter here beyond the headline number. First, the benefit is *distributionally* aimed at novices — the exact population the surge produces. Second, **retention improved**, in a domain whose turnover problem rhymes with direct care's.

The care-sector analogue already has evidence too, on the human side: structured transition-to-practice programs move first-year retention substantially. Reported baseline first-year turnover for new nurses runs about **31.7 percent**, while access to transition-to-practice programming is associated with turnover roughly **53 percent lower**, and mature system programs report aggregate first-year retention near **91 percent** ([Wolters Kluwer synthesis](https://www.wolterskluwer.com/en/expert-insights/why-nurse-residency-programs-improve-retention); [CommonSpirit Health](https://www.commonspirit.org/news-articles/nurse-residency-program-achieves-91-percent-retention-rate)). The intervention that works on novice retention is *structured supervised formation*. That is the thing to instrument.

### 4.2 The case against the naive version

Three findings, taken together, rule out the obvious design — a friendly chat assistant that answers a new care worker's clinical questions.

**(a) Model capability does not transfer through a novice user.** In a randomized trial with **1,298 UK participants**, the LLMs, tested alone on the same scenarios, identified relevant conditions in **94.9 percent** of cases and correct disposition in **56.3 percent**. The *same models used by the participants* yielded relevant conditions in **under 34.5 percent** of cases and disposition under **44.2 percent** — **no better than the control group** who used ordinary methods. Users did not know what to tell the model; the model's answers varied with small changes in phrasing; and responses mixed good and bad advice in ways participants could not separate ([Bean, Payne et al., *Nature Medicine*, 9 February 2026](https://www.ox.ac.uk/news/2026-02-10-new-study-warns-risks-ai-chatbots-giving-medical-advice); preprint [arXiv:2504.18919](https://arxiv.org/pdf/2504.18919)). Benchmarks measured the model. The trial measured the pair. The pair failed.

> Some secondary coverage of this study reports that the control group was "76 percent more likely" to identify correct conditions. The published framing this record relies on is the weaker, source-backed one: participants using LLMs performed **no better than** the control group.

**(b) Assistance degrades unassisted skill.** Across four endoscopy centres, the adenoma detection rate **for colonoscopies performed without AI fell from 28.4 percent before AI was introduced to 22.4 percent afterwards** — a 6.0 percentage-point absolute, roughly 20 percent relative, decline. It is the first real-world evidence of automation-induced deskilling tied to a patient-relevant outcome ([*The Lancet Gastroenterology & Hepatology*, August 2025](https://www.thelancet.com/journals/langas/article/PIIS2468-1253(25)00133-5/abstract)). Experienced clinicians lost capability from routine exposure. A workforce whose baseline skill is *still forming* is more exposed to this, not less.

**(c) Wrong advice actively flips right answers.** In a controlled prescribing study, **120 participants** working through scenarios with correct, incorrect, or absent decision support made substantially more commission errors when given incorrect alerts — **65.8 percent more in low-complexity, 53.5 percent more with interruption, 51.7 percent more in high-complexity scenarios** ([Lyell et al., *BMC Medical Informatics and Decision Making*, 2017](https://pmc.ncbi.nlm.nih.gov/articles/PMC5356416/)). Automation bias is not a theoretical risk; it is a measured effect whose magnitude scales with how prominently the wrong answer is displayed.

### 4.3 What that forces

C3 survives, in this form:

> Build novice-first — but the novice-facing product is **not an answer engine**. Its job is to orient, to hold scope, to make the user commit before it compares, to route to a named human fast, and to make supervision visible and cheap. Every place where the naive design would *answer*, this design *escalates or teaches*.

This is not a hedge invented for safety optics. It is what the three findings jointly require: (a) the pairing is the failure point, so design the pairing; (b) assistance erodes unassisted skill, so preserve unassisted practice deliberately; (c) confident wrong output flips correct human judgment, so never present clinical output prominently to someone without the scope to challenge it.

It also lines up with formation doctrine this project already holds — commit-then-compare learning, human evaluative authority, AI never grading alone ([`nurse-formation/DOCTRINE.md`](../nurse-formation/DOCTRINE.md)).

---

## 5. Evidence summary

| Finding | Figure | Source | Load it carries |
|---|---|---|---|
| Largest occupational job growth in the economy | +739,800 (17.0%), ~1 in 7 of all jobs added, 2024–34 | BLS MLR | C1 demand |
| Sector share of all new jobs | ~2.0M of 5.2M total, 2024–34 | BLS MLR | C1 demand |
| Near-term concentration | +680,500 (2.9%) Mar 2025→Mar 2026 while total nonfarm "changed little on net" | BLS TED | C1 demand |
| Total direct-care openings incl. exits | ~9.7M, 2024–34 | PHI Key Facts 2025 | C1 flow size |
| Pay | $17.36/hr median; under $26,000/yr median earnings (2024) | PHI Key Facts 2025 | C1 churn |
| Turnover | ~75% home care (2024); near 100% nursing-home aides (2017–18) | PHI | C1 churn |
| Immigrant share | 28% of direct care; 30% of aides; 32–40% home care | KFF | C1 fragility |
| Licensed-lane capacity | 93,176 qualified applications turned away (2025–26); 8.8% faculty vacancy; >60% of programs blocked on clinical placements | AACN data | C1 constraint |
| Supervisor supply | 39.9% RN / 41.3% LPN intent to leave in 5 years; median age 50 | NCSBN 2024 survey | The core constraint |
| Training floor | 75 hrs federal (≥16 supervised); 30 states + DC exceed; ~20 states unchanged ~30 yrs | 42 CFR 483.152; PHI | C2 variance |
| Title protection | DE HB 191; OR HB 2748; CA AB 489; WA HB 2155 | Holland & Knight; Nurse.org tracker | C2 boundary |
| Novice gains from AI | +15% avg; gains concentrated in less-experienced; retention up | Brynjolfsson, Li & Raymond, QJE 2025 | C3 support |
| Human-in-loop failure | Models alone 94.9% conditions; users with same models <34.5%, no better than control | Bean, Payne et al., *Nature Medicine* 2026 | C3 constraint |
| Deskilling | Non-AI ADR 28.4% → 22.4% after AI introduction | *Lancet Gastro Hepatol* 2025 | C3 constraint |
| Automation bias | +51.7% to +65.8% commission errors under incorrect decision support | Lyell et al. 2017 | C3 constraint |
| Formation works | ~31.7% first-year turnover baseline; ~53% lower with transition-to-practice access | Nurse residency literature | Design target |

---

## 6. What would falsify this

Stated in advance, so the strategy can be abandoned honestly rather than defended reflexively.

1. **The flow reverses.** Two consecutive BLS projection rounds showing health-care-and-social-assistance job growth falling below the all-industry average would remove the demand premise.
2. **Tenure stabilizes.** Direct-care turnover falling durably below ~30 percent — through wage floors, benefits, or policy — would mean the supervision bottleneck was a pay problem all along, and the software premise weakens accordingly.
3. **Supervision stops being scarce.** If experienced-nurse intent-to-leave drops materially and faculty and clinical-placement capacity expand, the constraint moves elsewhere and this strategy is aimed at the wrong thing.
4. **Novice assistance is shown to harm.** A trial of a scoped, escalation-first care-worker assistant showing worse unassisted performance, worse escalation timeliness, or more scope violations than an unassisted control ends this line of work regardless of adoption. Note the direction of the burden: the existing evidence in §4.2 is already unfavourable to naive designs, so this project must produce the evidence that its constrained design behaves differently. Absence of evidence is not a defence.
5. **Regulation forecloses it.** If the emerging state regime treats scope-and-escalation guidance to unlicensed care workers as clinical decision support requiring authorization this project does not hold, the personal edition must stop, not reinterpret itself.

## 7. Confidence and limits of this record

- All figures are from the cited sources; none are estimated, interpolated, or reconstructed from memory. Where a widely repeated secondary figure could not be confirmed in the primary source, the weaker confirmed version is used and the discrepancy is flagged (see §4.2a).
- U.S.-weighted. WHO data is the only non-U.S. input; conclusions about training floors, title protection, and employment law are U.S.-specific and do not transfer.
- Provenance is uneven and is stated rather than smoothed over. BLS employment figures and the federal nurse aide training rule are cited directly from the primary source. The nursing-workforce figures come from NCSBN's published survey and WHO's report — professional-body and agency primary sources. Two items rest on secondary reporting and are weaker for it: **the AACN capacity figures (93,176 applications turned away, 8.8 percent faculty vacancy, the clinical-placement share) are cited here through trade coverage of AACN's survey rather than the survey itself**, and **the state title-protection laws are cited through a law-firm review and a legislative tracker rather than the enacted texts** — only Delaware HB 191 is described from the legal review directly. Anyone relying on those two for a decision should read the underlying sources first. PHI is an advocacy organization publishing its own analysis of federal data; its figures are used because no equivalent aggregation exists, and they are labeled throughout.
- No causal claim is made linking AI displacement to caregiving entry. That link is plausible and unmeasured.
- This record is dated 2026-08-19 and will decay. The BLS projections round, the NCSBN survey cycle, and the state legislative sessions are the three clocks that matter.