All projections on AI Labs derive from peer-reviewed economics research and primary institutional studies. This page documents every source used, describes our forecasting model, and explicitly notes where evidence is strong, contested, or extrapolated.
We do not cite media analysis or opinion as evidence. Where institutional research (e.g. Goldman Sachs) is used, we cite the primary research report, not secondary coverage of it.
WHAT THE DATA ACTUALLY SHOWS AS OF EARLY 2026
The headline figure — that AI has fully replaced roughly 0.5–1% of US desk jobs — is a AI Labs synthesis, not a single study's finding. No single study measures "full replacement" directly. What the research measures are: task exposure, hiring deceleration, skill-demand shifts, and economic viability thresholds. We combine these to estimate actual displacement.
The honest answer is that full-replacement data is hard to isolate from other factors (post-pandemic normalization, interest rate effects on tech hiring, sector-specific cycles). Our 0.5–1% estimate is conservative and defensible, but it carries meaningful uncertainty.
The strongest observable signals of displacement to date: hiring growth has turned negative in marketing consulting, graphic design, office administration, and call centers. Technology-sector employment share has fallen below its pre-pandemic linear trend. These are consistent with AI-driven hiring deceleration, though not definitively causal.
FOUNDATIONAL THEORETICAL FRAMEWORK FOR ALL DISPLACEMENT MODELING
Establishes the displacement-reinstatement framework central to all serious AI labor modeling. Automation creates a displacement effect (machines replace labor in tasks) that reduces labor demand and wages, counteracted by a productivity/reinstatement effect (cost savings increase demand in non-automated tasks). The key empirical finding: these counteracting forces are incomplete — automation raises output per worker more than wages and reduces labor's share of national income.
Significance for our model: the framework predicts that displacement is not uniform across roles — it concentrates in tasks where AI has comparative advantage. This is why "task exposure" is the right unit of analysis, not "jobs."
NBER PAPER →Published empirical decomposition of US employment trends. Key finding: the slower growth of US employment over the last three decades is explained by an acceleration in the displacement effect (especially in manufacturing), a weaker reinstatement effect, and slower productivity growth than in previous technology waves. This has direct relevance to AI: unlike the PC revolution, which created substantial new tasks alongside automation, current AI deployment may not generate sufficient new task categories to offset displacement.
AEA PUBLICATION →Peer-reviewed in Econometrica — the field's most rigorous journal. Provides empirical general equilibrium estimates of automation's labor market effects. Confirms displacement effect has dominated reinstatement in recent decades, contributing to rising wage inequality. Used as the calibration basis for our task-substitution parameters.
NBER PAPER →ECONOMIC VIABILITY THRESHOLD — NOVEMBER 2025
The most important recent empirical contribution to this field. Unlike earlier "exposure" studies that measured theoretical task overlap between AI capabilities and job requirements, the Iceberg Index measures a stricter threshold: jobs where AI can perform the same tasks at a cost competitive with or lower than human labor. This is the economically relevant question — not what AI can do in principle, but where deployment is financially rational.
Key findings: AI can currently substitute for labor in roles representing 11.7% of the US labor market — approximately 151 million worker-equivalents — with roughly $1.2 trillion in annual wage value. The visible displacement (tech, IT roles) represents only $211 billion (2.2%); the remaining $1 trillion lies in HR, logistics, finance, and office administration — sectors not yet widely discussed in displacement coverage.
The study explicitly notes that capability does not automatically produce displacement — adoption speed, organizational inertia, and regulatory factors all mediate the conversion of technical capability into actual job loss. This is the basis for our "friction parameter" in the forecasting model.
The Iceberg Index is the primary source for our current-state estimate. The 11.7% figure represents economic viability today, not actual displacement. We estimate actual displacement at 0.5–1%, reflecting early-stage adoption against a technically feasible ceiling of ~12%.
SCHMIDT ET AL. — EMPLOYMENT EFFECTS AT THE FIRM LEVEL
Uses natural language processing to analyze ~58 million LinkedIn profiles and ~14 million job postings against 20,000 work activities in the DOL's O*NET database. Key finding: AI adoption at firms is associated with a 3.5% employment decline over five years in top-paying roles (management analysts, engineers, research scientists) — not in lower-paying roles as earlier technology waves affected. Business, financial, architecture, and engineering jobs shrank by 2–2.5% at AI-adopting firms.
Critical nuance: firms adopting AI do not necessarily shed workers — they grow and use workers more efficiently, meaning AI-exposed workers shift toward non-automated tasks within the same firm. This is the task reallocation mechanism that partially offsets the displacement effect at the individual firm level, even as aggregate hiring in affected roles slows.
NBER WORKING PAPER →QUANTITATIVE DISPLACEMENT ESTIMATES — BRIGGS & DONG (2025)
Analyzed more than 800 occupations to estimate displacement risk. Baseline finding: 6–7% of the US workforce faces displacement under wide AI adoption, with a range of 3–14% depending on adoption assumptions. Under current (limited) adoption scenarios, 2.5% of employment is at immediate risk from efficiency gains alone.
The research also estimates that generative AI could raise US labor productivity by ~15% when fully adopted — the productivity effect that partially offsets displacement at the macroeconomic level. Temporary unemployment during transition estimated at +0.5 percentage points above trend, historically dissipating within two years.
Highest-risk occupations identified: computer programmers, accountants and auditors, legal and administrative assistants, customer service representatives, telemarketers, proofreaders and copy editors, credit analysts. Lowest-risk: air traffic controllers, chief executives, radiologists, pharmacists, clergy.
Goldman's 6–7% baseline is explicitly described by the authors as a scenario under wide adoption, not a near-term forecast. It also assumes productivity gains partially offset displacement — a significant moderating assumption. Our more aggressive projections diverge from Goldman's because we incorporate recursive acceleration, which Goldman's model does not address.
GENERATIVE AI EFFECTS ON SKILL REQUIREMENTS — 2019–2024
Analyzes O*NET occupational data and US job postings from LightCast covering 2019–June 2024 across 923 occupations. Uses GPT-4o to calculate each occupation's generative AI exposure score, validated against human expert evaluations.
Key finding: following the introduction of generative AI, jobs in the top quartile of automation exposure showed a 24% decrease in generative AI-exposed skills per firm per quarter — indicating that these skills are being absorbed by AI. Jobs most susceptible to augmentation showed a 15% increase in AI-exposed skills, as workers develop complementary capabilities.
This is direct empirical evidence that GenAI is already reshaping what skills firms require, before the displacement is visible in employment statistics. It is consistent with the "stealth automation" thesis — structural changes happening below the surface of aggregate employment data.
HBS WORKING PAPER →EVIDENCE BASE AND HONEST LIMITS OF THIS CLAIM
The recursive acceleration thesis — that AI writing its own code shortens the doubling period, converting exponential growth into hyperbolic growth — is the most consequential and least well-grounded claim in our forecasting model. We are explicit about this.
What is directly evidenced: Anthropic has publicly stated that Claude writes approximately 90% of its own codebase. AI is being used in Neural Architecture Search and training pipeline optimization. These are documented facts, not projections.
What is extrapolated: The claim that this produces a self-reinforcing shortening of capability doubling periods is a plausible inference, not a measured empirical result. No peer-reviewed study has quantified the degree to which AI self-improvement compresses development timelines. The historical doubling estimates we cite (18 months → 10 months → 5 months) are AI Labs projections based on observed trends, not published research findings.
The key uncertainty: Recursive improvement could hit capability ceilings (architectural limits, training data constraints, compute bottlenecks) that prevent the hyperbolic curve from materializing. Energy infrastructure constraints are a genuine physical brake. The singularity framing is a mathematical extrapolation of a trend — it describes what happens if no ceiling is hit, not a prediction that no ceiling exists.
| CLAIM | EVIDENCE QUALITY | SOURCE TYPE |
|---|---|---|
| AI writes ~90% of its own code | Strong — direct statement | Anthropic (primary) |
| AI used in model architecture search | Strong — documented practice | Published ML research |
| Doubling period is shrinking | Moderate — trend extrapolation | AI Labs synthesis |
| Curve becomes hyperbolic by 2027–28 | Speculative — no published model | AI Labs projection |
| 50% displacement by 2029 | Speculative — scenario only | AI Labs scenario model |
INPUTS, PARAMETERS, AND STRUCTURAL ASSUMPTIONS
Our displacement projections use a modified adoption-curve model with three adjustable parameters. The interactive tool on this site allows readers to explore the output under their own assumptions.
| PARAMETER | BASE CASE VALUE | SOURCE / BASIS |
|---|---|---|
| Current displacement (2025) | ~1% | MIT Iceberg + hiring deceleration data |
| Economic viability ceiling (today) | 11.7% | MIT Iceberg Index, Nov 2025 (S04) |
| Wide-adoption displacement ceiling | 6–14% | Goldman Sachs Research (S06) |
| Base annual adoption speed | 2.5%/yr | AI Labs estimate from HBS skill data |
| Recursive multiplier (base case) | 1.35× | Extrapolated from Anthropic code-authorship trend |
| Friction dampener | 0.1 (minimal) | Corporate competitive dynamics assumption |
The base case produces 50% desk job displacement by approximately 2031–2033. The recursive scenario (multiplier 1.6, friction near zero) produces the 2029 estimate. The conservative scenario (multiplier 1.1, friction 0.6) produces 2040–2045.
HONEST UNCERTAINTY — THE LIMITS OF ALL CURRENT FORECASTING
No economist, AI researcher, or institution has successfully predicted the pace of AI capability growth more than 18 months in advance. Every major forecast from 2020–2023 has been revised — most dramatically upward — as capability improvements accelerated beyond model assumptions.
The fundamental problem is that standard economic forecasting assumes capability is an external input. In a recursive self-improvement regime, capability is endogenous. No published model adequately handles this. We include it as a scenario, not as a base case, for this reason.
Additional unknowns: the degree to which productivity gains raise wages and create new demand (potentially offsetting displacement); the speed at which new task categories emerge from AI; the possibility of a capability plateau; policy responses including regulation, UBI experiments, or mandated hiring floors; and the feedback effects of mass unemployment on consumer demand and corporate investment.
Treat all projections on this site as scenario maps, not predictions. The value is in understanding the structure of the uncertainty, not the precise year.