---
id: wiki-2026-0508-work-displacement
title: Work Displacement (AI 노동 대체)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Job Displacement, AI Automation, Labor Displacement, Technological Unemployment]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [labor, economics, ai-impact, automation, policy]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python/R
  framework: pandas/statsmodels
---

# Work Displacement (AI 노동 대체)

## 매 한 줄
> **"매 AI/automation 이 매 specific tasks/occupations 의 demand 를 매 reduce/eliminate — 매 net unemployment 와 매 different concept (재배치 가능성 포함)"**. Frey & Osborne 2013 47% 추정 부터 매 IMF 2024 (40% global jobs 영향), Goldman Sachs 2023 (300M FTE), MIT Acemoglu 2024 의 매 conservative estimate (5-10% GDP 효과 / 10년) 까지 매 다양한 추정. 2026 현재 매 white-collar (legal, code, design, customer support) 가 매 first-wave 에 있고, 매 재훈련 / wage 압력 / inequality 가 매 policy 핵심 의제.

## 매 핵심

### 매 displacement vs replacement vs augmentation
- **Displacement**: task 자체 매 사라짐 (e.g., elevator operator).
- **Replacement (substitution)**: human 매 machine 으로 대체.
- **Augmentation**: human + AI 매 productive — 매 어떤 worker 매 더 valuable.
- **Compensation effects**: 매 displaced worker 매 new task / lower-cost product / new sector 로 흡수 (역사적 trend).

### 매 task 단위 분석 (Acemoglu-Restrepo framework)
- 매 occupation = bundle of tasks. AI 매 일부 task 만 자동화 — 매 occupation 자체 disappear 보다 매 task composition 변경 보통.
- 매 "exposure" ≠ "automation" — 매 GPT 가 task 처리 가능 means exposed, but adoption / cost / regulation 매 lag.

### 매 affected categories (2026 evidence)
- **Highly exposed**: paralegal, copywriter, junior translator, basic illustrator, tier-1 customer support, junior dev (boilerplate code), tax prep.
- **Augmented**: senior eng, doctor (radiology + AI co-read), researcher, analyst, designer.
- **Insulated**: hands-on physical (plumber, surgeon), high-touch interpersonal (therapy, teaching young kids), strategic / political.

### 매 empirical findings (2024-2026)
- Brynjolfsson et al. 2023: customer support — 매 14% productivity, 매 novice gains > expert.
- MIT writing study 2023: 37% time reduction, quality up.
- Goldman 2024: 매 entry-level coder demand 매 5-15% 감소 signal (US BLS data).
- WEF 2025 Future of Jobs: 매 net 7M jobs decline 2025-2030, 매 AI-adjacent role 매 10M 증가.

### 매 policy responses
1. UBI / negative income tax — 매 still pilot stage.
2. Reskilling / displacement insurance (Singapore SkillsFuture, EU Just Transition).
3. Working time reduction (4-day week trials).
4. AI-specific tax (Bill Gates proposal).
5. Sectoral protection (e.g., truckers vs autonomous trucks).

### 매 응용
1. Workforce planning (어느 role 매 AI exposure 높은지).
2. Curriculum design (대학, bootcamp).
3. Public policy modeling.
4. Corporate L&D investment 결정.
5. Personal career strategy.

## 💻 패턴

### 1. O*NET task exposure score (Python)
```python
# Compute occupation-level AI exposure from task descriptions
import pandas as pd
from sentence_transformers import SentenceTransformer
import numpy as np

embedder = SentenceTransformer("BAAI/bge-large-en-v1.5")

# AI capability descriptors (cf. Eloundou et al. 2023 GPTs are GPTs)
AI_CAPS = [
    "summarize a long document",
    "draft an email or report",
    "write or debug code",
    "extract structured data from text",
    "answer factual questions",
    "translate between languages",
    "generate an image from a description",
]

tasks = pd.read_csv("onet_tasks.csv")  # task_id, occupation, task_text
ai_emb = embedder.encode(AI_CAPS, normalize_embeddings=True)

def exposure(task_text: str) -> float:
    t = embedder.encode(task_text, normalize_embeddings=True)
    return float(np.max(t @ ai_emb.T))

tasks["exposure"] = tasks.task_text.apply(exposure)
occ = tasks.groupby("occupation").exposure.mean().sort_values(ascending=False)
print(occ.head(20))
```

### 2. Wage impact regression
```python
import statsmodels.formula.api as smf
import pandas as pd

# Cross-sectional wage / employment by occupation × AI exposure
df = pd.read_csv("bls_acs_2025.csv")  # occ, exposure, wage_change_pct, employment_change_pct
m = smf.ols("wage_change_pct ~ exposure + edu_years + female_share + remote_share", data=df).fit()
print(m.summary())
```

### 3. Difference-in-differences (ChatGPT shock 2022→2025)
```python
# Treatment: high-exposure occupations, post: 2023+
import statsmodels.formula.api as smf

panel["post"] = (panel.year >= 2023).astype(int)
panel["treat"] = (panel.exposure > panel.exposure.median()).astype(int)
panel["did"] = panel.post * panel.treat

m = smf.ols(
    "log_employment ~ post + treat + did + C(occ) + C(year)",
    data=panel,
).fit(cov_type="cluster", cov_kwds={"groups": panel.occ})
print(m.params["did"], m.pvalues["did"])
```

### 4. Reskilling pathway recommender
```python
# For displaced worker, suggest adjacent occupations with low exposure
def recommend(current_occ: str, df: pd.DataFrame, top_k=5):
    cur_skills = df[df.occ == current_occ].iloc[0].skill_vector  # pre-computed
    candidates = df[df.exposure < df.exposure.quantile(0.3)]
    candidates = candidates.assign(
        sim=candidates.skill_vector.apply(lambda v: float(v @ cur_skills)),
    )
    return candidates.nlargest(top_k, "sim")[["occ", "wage_median", "sim", "exposure"]]
```

### 5. Macro impact simulation (production-function)
```python
# Cobb-Douglas with AI as capital-like factor
import numpy as np
def gdp(L, K, AI, alpha=0.3, beta=0.5, gamma=0.2, A=1.0):
    return A * L**alpha * K**beta * AI**gamma

# Scenario: AI capital grows 30%/yr, labor productivity 비례
years = np.arange(2026, 2036)
L0, K0, AI0 = 100, 100, 10
gdp_path = [gdp(L0 * 1.005**i, K0 * 1.02**i, AI0 * 1.30**i) for i in range(len(years))]
print({y: round(v, 1) for y, v in zip(years, gdp_path)})
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Workforce planning corporate | O*NET exposure + internal task mining → role redesign |
| Personal career planning | high-touch / strategic / physical tasks 비중 증가 방향 |
| Public policy | sectoral exposure 분석 + reskilling / safety net 설계 |
| L&D investment | augmentation skill (prompt eng, AI literacy) > 단순 tool training |
| Hiring strategy | senior + AI-augmented 매 leverage 매 큼; junior pipeline 재설계 |

**기본값**: 매 task 단위 분석 (occupation 단위 X) → 매 augmentation-first framing (replacement-first 매 over-pessimist).

## 🔗 Graph
- 부모: [[Technological Unemployment]]
- 응용: [[UBI]]

## 🤖 LLM 활용
**언제**: 매 strategic workforce/career decision, 매 policy analysis, 매 corporate restructure planning.
**언제 X**: 매 short-term predictions about specific job count 매 unreliable — 매 directional / probabilistic framing.

## ❌ 안티패턴
- **Lump of labor fallacy**: 매 work 의 fixed pie 가정 — 매 historical 으로 매 false. 매 new task / new sector 매 emerge.
- **Occupation-level only**: 매 task-level 이 정확한 단위.
- **Frey-Osborne 47% 매 prediction 으로 인용**: 매 exposure score 매 actual displacement 와 다름.
- **AI-doomer + AI-utopian extremes**: 매 둘 다 evidence 와 매 잘 맞지 않음 — 매 transition cost 와 매 distributional impact 가 핵심.
- **Reskilling = silver bullet**: 매 mid-career reskilling 매 historically low success rate (~30%) — 매 supplemented by social safety net 필요.

## 🧪 검증 / 중복
- Verified (Acemoglu & Restrepo 2018, Eloundou et al. 2023, IMF Gen-AI Working Paper 2024, WEF Future of Jobs 2025).
- 신뢰도 A (empirical), B (forecasts).

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — task-level framework + empirical findings + policy options |