証拠の質 4.88/5
8次元のレビュー評価。基準は 品質ルーブリック 。各次元は1〜5で評価。
- D1 出典への根拠
- 5/5
- D2 出典の権威性
- 5/5
- D3 算術
- 5/5
- D4 不確実性
- 4/5
- D5 範囲
- 5/5
- D6 文章
- 5/5
- D7 認識の誠実性
- 5/5
- D8 注意事項の完全性
- 5/5
リスクはどう変わるか
見出しの数値は大きく異なる状況の平均です。シナリオや文脈による確率の違いは以下の通りです:
2.0分の1 · 50%
Highest-exposure category across all studies; Frey & Osborne assign >90% automation probability to many of these roles. But 'high exposure' ≠ 'job eliminated' — partial automation is far more common than full replacement.
5.0分の1 · 20%
Moderate task exposure per Eloundou et al., but augmentation (AI as tool) is the dominant pattern so far, not replacement. Historical analogies: spreadsheets did not eliminate accountants.
6.7分の1 · 15%
Generative AI has high task exposure here, but commercial adoption for full replacement remains low. Copyright and quality-control constraints limit substitution.
20分の1 · 5.0%
Lowest AI exposure. Robotics advances lag software AI by a wide margin. These roles also face strong demographic tailwinds (aging population, labor shortages).
バーの長さと濃淡は、これらのシナリオを他のリスクとではなく、互いに比較して順位付けしたものです。正確な確率は各項目の横に表示されています。
このリスクが関わる決断
このリスク次第で変わる選択。そのトレードオフをどう見るか。
「AIがあなたの仕事を奪う確率はどのくらいか?」という問いに答える単一の数字は存在しません。なぜなら、そのような研究結果はまだ発表されていないからです。最も引用される数字であるFreyとOsborneの米国雇用の47%が高リスクという数値は、失業の予測ではなく、自動化に対するタスクレベルの感受性を表しています。OECDが職種レベルの集計データではなく個人レベルのデータを用いて分析を再実行したところ、高リスクの割合は**9~14%**に低下しました。2023年のGPTs-are-GPTs論文では、労働者の80%がLLMに何らかのタスクで曝露しているものの、タスクの半分以上が曝露しているのは19%に過ぎないと報告されています。一方、労働統計局は2033年までに670万人の純雇用増加を予測しています。これらの数字は矛盾しているわけではなく、それぞれ異なるものを測定しているのです。
「AIが実行できるタスク」と「消滅する仕事」との間のギャップに、ほとんどの不確実性が存在します。ATMは銀行の窓口係をなくすはずでしたが、実際には支店の運営コストが下がったことで支店数が増え、窓口係の雇用は20年間増加しました。生成AIが同じような拡張パターンをたどるのか、それともそれを打ち破るのかが中心的な実証問題であり、2026年初頭時点での正直な答えは、誰にも分からないということです。研究は曝露の方向性については一致しています。定型的な認知作業(データ入力、基本的な文書レビュー、簿記など)は最も高い代替圧力を受ける一方、手作業・身体作業(建設、配管、高齢者介護など)は最も低い圧力を受けています。しかし、曝露は運命ではありません。
個人のリスクは、職種、産業、地域によって少なくとも桁違いに異なります。データ入力担当者と配管工は異なるリスクの世界に生きており、どちらも人口平均で適切に説明されるものではありません。時間軸も同様に重要です。5年間の見通しは拡張が支配的ですが、30年間の見通しは本当に不確実です。発表されている推定値の幅、これまでの大規模な代替の不在、そしてタスク曝露と仕事の消滅との間の根本的なギャップを考慮すると、この項目にはno_reliable_estimateのフラグが付いています。恐怖は現実のものですが、擁護できる生涯確率は存在しません。
根拠台帳
以下の各数値は各出典が報告した内容であり、引用した原文の抜粋と算出方法を記載しています。リンクをクリックして直接確認できます。
-
[1] Oxford Martin School, University of Oxford — The Future of Employment: How Susceptible Are Jobs to Computerisation?
The Future of Employment: How Susceptible Are Jobs to Computerisation?- 統計値
47% of US employment is in the high-risk category for computerisation over the next decade or two- 抜粋
“"We examine how susceptible jobs are to computerisation. … According to our estimate, 47 percent of total US employment is in the high risk category, meaning that associated occupations are potentially automatable over some unspecified number of years, perhaps a decade or two." ”
- 出典データ
- 2013-09-17
- アクセス日
- 2026-04-18 · アーカイブ版
- 計算過程
- Frey & Osborne (2017, published in Technological Forecasting and Social Change 114:254-280) classified 702 occupations by automation probability using O*NET task descriptors. The 47% figure refers to jobs whose constituent tasks are technically automatable, not a prediction that 47% of workers will be unemployed. The authors explicitly note this is a measure of task susceptibility, not a forecast of job losses. The paper predates the LLM era and does not account for generative AI capabilities. No conversion to a per-worker lifetime displacement probability is attempted because the study does not model job creation, reallocation, or partial automation.
- 独立性
- Seminal academic study using O*NET occupational data. Independent methodology from the OECD and Goldman Sachs estimates below.
-
[2] OECD — The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis
The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis- 統計値
9% of jobs across the 21 OECD countries are highly automatable; an additional ~25% face significant task change- 抜粋
“"Overall, we find that, on average across the 21 OECD countries, 9 % of jobs are automatable. [...] For instance, while the share of automatable jobs is 6 % in Korea, the corresponding share is 12 % in Austria." ”
- 出典データ
- 2016-05-14
- アクセス日
- 2026-04-18 · アーカイブ版
- 計算過程
- Arntz, Gregory & Zierahn (2016) re-analysed Frey & Osborne's methodology using PIAAC individual-level survey data rather than occupation-level aggregates. By accounting for within-occupation task variation, the high-risk share drops from 47% to ~9-14%. The OECD updated this to 14% in its 2019 Employment Outlook using newer PIAAC data. Neither figure is a probability of displacement for an individual worker; it describes the share of jobs with >70% automatable task content. No lifetime probability is derivable from this study because it does not model transition rates or job creation.
- 独立性
- Uses PIAAC microdata, a different data source and unit of analysis (individual workers vs occupations) from Frey & Osborne. Methodologically independent.
-
[3] OpenAI / University of Pennsylvania — GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models
GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models- 統計値
~80% of the US workforce could have at least 10% of their tasks affected by LLMs; ~19% could have 50%+ of tasks affected- 抜粋
“"Our findings reveal that around 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs, while approximately 19% of workers may see at least 50% of their tasks impacted." ”
- 出典データ
- 2023-03-17
- アクセス日
- 2026-04-18 · アーカイブ版
- 計算過程
- Eloundou, Manning, Mishkin & Rock (2023) assessed task-level exposure to LLMs across 1,016 occupations using O*NET. "Exposure" means the task could be completed significantly faster with LLM access, not that the job would be eliminated. The 80% headline describes partial task exposure (>=10% of tasks), not full job replacement. The 19% figure for >=50% task exposure is closer to a displacement signal but still measures task content, not employment outcomes. Converting task exposure to a lifetime job-loss probability requires assumptions about employer adoption rates, augmentation vs replacement decisions, and labor market adjustment speeds that the paper does not provide.
- 独立性
- Uses O*NET task data (shared with Frey & Osborne) but applies a GPT-specific exposure rubric developed by both human annotators and GPT-4 classification. Partially independent.
-
[4] Goldman Sachs — The Potentially Large Effects of Artificial Intelligence on Economic Growth
The Potentially Large Effects of Artificial Intelligence on Economic Growth- 統計値
Two-thirds of US occupations are exposed to some degree of AI automation; generative AI could substitute up to 25% of current work tasks globally- 抜粋
“"our economists estimate that roughly two-thirds of U.S. occupations are exposed to some degree of automation by AI. They further estimate that, of those occupations that are exposed, roughly a quarter to as much as half of their workload could be replaced." ”
- 出典データ
- 2023-03-26
- アクセス日
- 2026-04-18 · アーカイブ版
- 計算過程
- Hatzius et al. (2023) estimate that generative AI could expose 300 million full-time jobs globally to automation, with roughly 25% of work tasks substitutable. The report simultaneously projects a 7% increase in global GDP over a 10-year period from AI adoption, implying substantial job creation alongside displacement. The 300M figure describes exposure, not net job loss, and the report explicitly models offsetting job creation from new industries and increased productivity. No per-worker lifetime displacement probability can be extracted.
- 独立性
- Goldman Sachs proprietary analysis drawing on O*NET, European ISCO data, and internal economic modelling. Partially independent from academic studies above.
-
[5] Bureau of Labor Statistics — Employment Projections: 2023-2033
Employment Projections: 2023-2033- 統計値
Total US employment projected to grow by 6.7 million jobs from 2023 to 2033, to 174.6 million (0.4 percent a year)- 抜粋
“"The U.S. economy is projected to add 6.7 million jobs from 2023 to 2033, the U.S. Bureau of Labor Statistics (BLS) reported today. Total employment is projected to increase to 174.6 million and grow 0.4 percent annually, which is slower than the 1.3 percent annual growth recorded over the 2013-23 decade." ”
- 出典データ
- 2024-09-04
- アクセス日
- 2026-04-18 · アーカイブ版
- 計算過程
- BLS decennial projections are the baseline measure of actual expected employment outcomes. The 2023-2033 outlook projects net job growth of 6.7 million, not net loss. BLS does not separately model AI-driven displacement; its projections incorporate historical trends in technology adoption, productivity growth, and sectoral shifts. The fact that net employment is projected to grow does not preclude significant churn within occupations, but it provides a counterweight to headline claims of mass job elimination. No per-worker lifetime displacement probability is published.
- 独立性
- Official US government statistical agency using proprietary macroeconomic modelling. Fully independent from the academic and private-sector studies above.
-
[6] Stanford Digital Economy Lab — Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence
Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence- 統計値
Early-career workers (ages 22-25) in the most AI-exposed occupations experienced a 16 percent relative decline in employment since late 2022, while employment for more experienced workers in the same occupations and for workers in less-exposed fields remained stable or continued to grow- 抜粋
“"early-career workers (ages 22-25) in the most AI-exposed occupations have experienced a 16 percent relative decline in employment even after controlling for firm-level shocks. In contrast, employment for workers in less exposed fields and more experienced workers in the same occupations has remained stable or continued to grow." ”
- 出典データ
- 2025-11-13
- アクセス日
- 2026-06-14 · アーカイブ版
- 計算過程
- Brynjolfsson, Chandar & Chen (2025) analyse high-frequency payroll records from ADP, the largest US payroll processor, covering millions of workers. They document a 16% relative decline in headcount for the 22-25 age band in occupations most exposed to generative AI (software developers, customer-service representatives) measured against pre-2023 trend, with no comparable decline for workers 30+ in the same occupations or for workers in low-exposure fields. The 16% is a cohort-level relative employment differential, not a per-worker probability of job loss; the paper does not convert it into an individual displacement hazard, and AI attribution remains contested (a Feb 2026 follow-up by the same authors examines interest-rate and timing confounds). It is used here only to populate the early-career age x AI-exposure personal factor, not to derive a headline number. The earlier August 2025 preprint reported ~13%; the published version reports 16%.
- 独立性
- First large-scale payroll-based (ADP administrative records) evidence, methodologically independent from the O*NET task-exposure studies and BLS macroeconomic projections above. Measures realised employment, not task susceptibility.
-
[7] Proskauer (Law and the Workplace) — AI Bias Lawsuit Against Workday Reaches Next Stage as Court Grants Conditional Certification of ADEA Claim
AI Bias Lawsuit Against Workday Reaches Next Stage as Court Grants Conditional Certification of ADEA Claim- 統計値
On May 16, 2025 a federal court conditionally certified Mobley v. Workday as a nationwide ADEA collective covering job applicants aged 40+; Workday's filings disclosed roughly 1.1 billion applications rejected through its AI screening tools during the relevant period- 抜粋
“"1.1 billion applications were rejected [...] Notice will be issued to the allegedly affected job applicants, in what could be one of the largest collectives ever certified." ”
- 出典データ
- 2025-06-03
- アクセス日
- 2026-06-14 · アーカイブ版
- 計算過程
- Mobley v. Workday (N.D. Cal. 3:23-cv-00770) alleges Workday's AI applicant-recommendation system produced a disparate impact by age, race, and disability at the hiring gate. On 2025-05-16 the court granted conditional (preliminary) certification of a nationwide ADEA collective of applicants aged 40+ screened since September 2020. The ~1.1 billion figure is the volume of applications Workday disclosed were rejected through its tools, not a number of confirmed biased rejections; the case is unresolved and the disparate-impact claim is untested at trial. No per-applicant probability of a bias-tainted auto-rejection is derivable, so this source supports a caveat only, not a number.
- 独立性
- Legal-developments analysis of court filings; non-authoritative reference used only to document the algorithmic-exclusion-at-the-hiring-gate dimension absent from the displacement-by-automation studies above.

