Anthropic 檢視 56 項隨機試驗：傳統再訓練成效有限，難以因應 AI 大規模衝擊
Reviewing the evidence on worker retraining programs
摘要分析
Anthropic 於 2026 年 8 月 12 日發布由 David Roodman 與 Maxim Massenkoff 合著的證據回顧，探討「再訓練」這項最受歡迎的 AI 衝擊因應政策是否奏效。統合分析 56 項美國隨機試驗（146 個效果估計）發現職訓效果「正面但有限」：每個名額讓就業率升 2–3 個百分點、年收入增約 1,000 美元，成本約 1.3 萬美元；計入稅收與減少的福利後大致打平。與高需求產業雇主合作、直接安置就業的產業別計畫效果可達數倍，但複製常失敗。結論：若 AI 造成大規模失業，現有再訓練恐不足，應現在就投資示範、評估並擴大最有潛力的模式。
- 56 項隨機試驗：職訓讓就業率僅升 2–3 個百分點
- 年收入增約 1,000 美元，每名額成本 1.3 萬美元
- 與雇主合作、直接安置的產業別計畫效果數倍
- 成功模式難以複製
- 現有再訓練難以因應 AI 大規模衝擊
Anthropic published an evidence review on August 12, 2026, co-authored by independent researcher David Roodman and Anthropic economist Maxim Massenkoff, asking whether retraining—the most popular policy response to AI-driven labor disruption—would work at scale. A new meta-analysis of 56 randomized US trials (146 impact estimates from 1973 onward), plus European evidence, finds job-training programs have positive but modest effects: for each person offered a slot, employment rises 2–3 percentage points and earnings by about $1,000 a year, at a cost of roughly $13,000 per slot. Counting added tax revenue and reduced benefits, governments recover more than half the spending, so programs roughly break even. Sectoral programs that partner with employers in high-demand industries and place people directly into jobs yield gains several times larger, but replication attempts often fail. The authors conclude existing programs would likely fall short if AI displaces workers at scale and recommend investing now to demonstrate, evaluate and scale the most promising models.
- 56 RCTs: training lifts employment only 2–3 points
- Earnings up ~$1,000/year at ~$13,000 per slot
- Employer-linked sectoral programs deliver several times the gains
- Successful models are hard to replicate
- Existing retraining would fall short under large-scale AI displacement
關鍵字：再訓練隨機試驗產業別訓練勞動政策Anthropic
來源原文：https://www.anthropic.com/research/reviewing-the-evidence-on-worker-retraining-programs
本頁為報導內容的摘要與分析，版權屬原媒體所有；請以原文為準。