Predictions Over Time

Only data points with directly sourced displacement figures are plotted. Directional studies (Brookings, Yale Budget Lab, Stanford/World Bank) appear as overlay markers. Exposure-based estimates (Goldman, OECD, IMF) have been moved to the separate Exposure & Risk tracker.

Exposure-based projection. Early evidence from HBS shows job postings in automatable professional roles declining 17%, but augmentation-prone professional roles growing 22% (Chen et al., 2025).

Key nuance: HBS job postings data shows a 17% decrease in highly automatable tech occupations but a 22% increase in augmentation-prone roles (Chen et al., 2025) — net displacement may be lower than gross.

Partially observed: Ramp spending data shows freelance marketplace spend (heavily creative) collapsed from 0.66% to 0.14% of total business spend. Upwork and Fiverr have reported significant volume declines in writing and design categories.

Exposure-based with early signals. The collapse of Chegg subscribers (~50%) after ChatGPT launch is an observed effect in adjacent education services, though core teaching roles remain among the least automatable.

Mostly exposure-based. Healthcare admin has high task automability scores but adoption has been slower than tech or creative sectors due to regulatory requirements and EHR integration challenges.

Mixed evidence: Citigroup estimates 54% of banking jobs have high automation potential, and Goldman Sachs surveys project 14% headcount reduction over 3 years. However, observed Big Six bank headcount fell only ~1% in 2025, and Yale/Fortune analysis calls current displacement 'smoke and mirrors.' Major banks (Citi, Morgan Stanley, BofA) are announcing cuts, but net displacement remains well below projections.

Supported: Stanford payroll data shows early-career CS worker declines similar to software developers (Brynjolfsson et al., 2025). One of few areas with both exposure estimates and observed substitution from field data.

Mixed evidence: Acemoglu & Restrepo (2020) provide causal estimates from US labor markets showing modest displacement per robot. IFR data shows accelerating installations but BLS projections remain conservative. Physical task automation is cost-constrained. McKinsey (2025) notes robot unit costs must fall 70-90% for mass adoption.

Mixed signals: del Rio-Chanona et al. (2025) find productivity gains of 20–60% in experiments, which should push wages up, but the Anthropic Economic Index shows automation (45%) competing with augmentation (52%) — the net wage effect depends on which pattern dominates.

Strong empirical support: Stanford payroll study found ~20% headcount decline for developers aged 22–25 (Brynjolfsson et al., 2025); Dallas Fed CPS data shows employment share of young workers in AI-exposed occupations fell from 16.4% to 15.5%; IMF finds employment 3.6% lower in AI-vulnerable occupations after 5 years, with young workers most exposed.

Complicated by new evidence: AI exposure measures converge toward high-wage jobs (del Rio-Chanona et al., 2025), but Brookings research found high-skill freelancers are disproportionately affected — not insulated (Hosseini & Lichtinger, 2025). HBS data shows 22% increase in augmentation-prone job postings, suggesting the premium may accrue to AI-complementary skills specifically.

Confirmed by multiple sources: Ramp firm spending data shows freelance marketplace spend fell from 0.66% to 0.14% of total business spend while AI provider spend rose to ~3%. Digital trace data confirms substitution in writing and translation (del Rio-Chanona et al., 2025).

Directly observed from Census Bureau Business Trends survey data, supplemented by corporate earnings call disclosures and Ramp firm spending analysis.

Directly observed from the St. Louis Fed / Harvard Real-Time Population Survey (RPS), a nationally representative quarterly survey of 25,000+ working-age adults. Feb & May 2025 values estimated from tracker chart; all other values from published working paper.

Directly from source studies. These measure what share of jobs have tasks that AI could perform — not how many jobs have actually been affected. Exposure ranges from 27% (OECD, high-risk only) to 80% (OpenAI, any task overlap). The Anthropic Economic Index provides the most current measure at 49% of jobs using AI for 25%+ of tasks.

Directly observed from NLP analysis of S&P 500 earnings call transcripts. Companies mentioning AI + workforce subsequently show headcount growth 3.2pp lower than non-mentioners.