Forecasting Methodology
& Analytical Framework

01. Nine-Variable Framework

THE STRUCTURAL INPUTS THAT DRIVE EVERY AI LABS PROJECTION

Every displacement forecast published on this site is generated from a model built on nine core variables. These variables were selected because they represent the minimum set of independent causal factors required to model AI labor displacement with structural fidelity. Each variable is measurable (though measurement quality varies), each has documented sources, and each can be independently adjusted to explore different futures.

The framework is deliberately parsimonious. More variables would increase apparent precision but not actual accuracy. We model the forces that matter most and are transparent about what is excluded.

02. Scenario Construction Methodology

HOW CONSERVATIVE, BASE CASE, AND RECURSIVE SCENARIOS ARE BUILT

All AI Labs projections are published as three parallel scenarios, not point predictions. Each scenario represents a coherent, internally consistent set of V1-V9 parameter values. The scenarios are not optimistic/pessimistic -- they correspond to different structural assumptions about the world.

The table below documents the exact parameter settings for each scenario. Readers can use the interactive displacement model on this site to explore intermediate parameter combinations.

Conservative scenario assumes AI capability growth decelerates to a 12-month doubling period (consistent with a paradigm plateau), self-improvement effects are negligible, adoption faces persistent organizational friction, and regulatory environments significantly constrain deployment speed. This scenario is consistent with the position held by most mainstream labor economists.

Base case scenario extends current METR-measured trends without acceleration or deceleration. The recursive self-improvement loop exists but is moderate (1.35x multiplier, derived from Anthropic's code-authorship data). Adoption friction is real but erodes under competitive pressure. This is our central projection and the one used in headline forecasts.

Recursive scenario assumes the self-improvement feedback loop fully materializes, producing hyperbolic rather than exponential growth. Adoption friction collapses as the cost differential becomes irresistible. Regulatory response is slow. This is a tail-risk scenario, not a central forecast -- but the METR acceleration trend (7 months overall, 4 months since 2023) means it cannot be dismissed.

03. Data Sources Overview

PRIMARY INPUTS TO THE FORECASTING MODEL

All AI Labs projections derive from primary institutional research and direct data sources. We do not cite media analysis, opinion columns, or secondary coverage as evidence. Where institutional research (e.g., Goldman Sachs) is used, we reference the primary research report, not news articles about it. Our full source documentation is available on the Sources page.

• ACADEMIC Acemoglu & Restrepo (NBER, 2018/2019/2022) -- the task-based displacement-reinstatement framework that provides the theoretical foundation for all serious AI labor modeling. Published in Econometrica (the field's most rigorous journal). Calibration basis for V6 and V7 parameters.
• ACADEMIC MIT Iceberg Index (CSAIL, November 2025) -- economic viability threshold study measuring where AI substitution is financially rational, not merely technically feasible. Primary source for V3 calibration and the current-state displacement estimate (11.7% viable, ~1% actual).
• ACADEMIC Harvard Business School Working Paper 25-039 -- firm-level analysis of generative AI effects on skill requirements using O*NET and LightCast data (923 occupations, 2019-2024). Documents 24% decline in AI-exposed skills at high-automation firms.
• ACADEMIC Schmidt et al. (NBER Working Paper 33509) -- 58 million LinkedIn profiles and 14 million job postings analyzed against O*NET activity data. Shows 3.5% employment decline in top-paying roles at AI-adopting firms over five years.
• BENCHMARKS METR Autonomous Task Benchmark (TH1.1) -- the primary calibration source for V1 (AI Capability Growth Rate). Measures real-world autonomous task completion by frontier models. 15 data points across 6 years. Current frontier: Claude Opus 4.6 at 14h 30m human-equivalent task complexity. Doubling period: ~7 months overall, ~4 months since 2023.
• GOVERNMENT Bureau of Labor Statistics (BLS) -- monthly employment data (Current Employment Statistics, CES), occupational employment and wage statistics (OES), quarterly census of employment and wages (QCEW). Used for baseline labor market parameters, wage data (V5 calibration), and real-time employment trend validation.
• GOVERNMENT 2026 International AI Safety Report -- cross-government assessment of AI capability trajectories and safety risks. Documents AI evaluation-awareness and behavioral modification capabilities relevant to V2 assessment.
• CORPORATE Goldman Sachs Global Investment Research (Briggs & Dong, 2025) -- 800+ occupation analysis estimating 6-7% workforce displacement under wide adoption (range 3-14%). Provides V7 calibration and sector-level risk decomposition.
• CORPORATE Quarterly earnings calls and investor presentations -- direct statements from technology companies, major employers, and AI infrastructure providers regarding AI deployment, headcount decisions, and cost savings. Sources include Anthropic, OpenAI, Google DeepMind, Microsoft, Block (Square), and others. Used for real-time V3 and V5 calibration.
• TRACKERS Technology layoff trackers (Layoffs.fyi, TrueUp, Challenger Gray) -- aggregated layoff announcements cross-referenced with company AI investment disclosures to identify AI-correlated workforce reductions versus cyclical adjustments. Used as a leading indicator, not as a primary data source, due to attribution uncertainty.
• REGULATORY EU AI Act, proposed US AI frameworks, sector-specific guidance (SEC, HHS, DOL) -- tracked for V4 (Regulatory Drag) calibration. Regulatory trajectory monitored for both direct deployment restrictions and indirect effects (compliance costs, liability exposure, mandatory human-in-the-loop requirements).

04. Confidence Calibration System

WHAT HIGH, MEDIUM, AND LOW CONFIDENCE MEAN -- AND WHAT EVIDENCE IS REQUIRED FOR EACH

Every projection, claim, and parameter estimate published on AI Labs carries a confidence designation. These are not subjective feelings -- they correspond to specific evidence thresholds. The system is designed to prevent a common failure mode in forecasting: presenting speculative extrapolations with the same rhetorical weight as empirically grounded measurements.

05. Update Cadence

WHEN AND HOW PROJECTIONS ARE REVISED

Forecasting models are only as good as their maintenance discipline. Static projections degrade rapidly in a domain where the underlying dynamics shift on monthly timescales. AI Labs maintains a structured revision schedule with three tiers.

All revisions include a change log documenting which parameters moved, in which direction, by how much, and why. Historical parameter values are preserved for audit and calibration tracking. We do not silently update projections -- every change is versioned and explained.

06. Caveats & Limitations

HONEST ASSESSMENT OF WHAT THIS MODEL DOES NOT AND CANNOT CAPTURE

No forecasting model is better than its weakest structural assumption. We are explicit about the limitations of this framework because intellectual honesty is more valuable than false precision. Users of these projections should weight them accordingly.

1. No model has successfully predicted AI capability growth more than 18 months in advance. Every major forecast from 2020-2023 required significant upward revision. Our model inherits this fundamental limitation. Projections beyond 2028 should be treated as scenario explorations, not timeline commitments.
2. The recursive self-improvement variable (V2) is the least empirically grounded parameter. While AI code self-authorship is documented, no peer-reviewed study has quantified the degree to which this compresses capability doubling periods. The hyperbolic curve geometry in the recursive scenario is a mathematical extrapolation of a plausible dynamic, not an observed phenomenon. It may encounter ceilings (architectural limits, training data constraints, compute/energy bottlenecks) that flatten the curve before hyperbolic dynamics manifest.
3. The model does not capture new task creation with structural fidelity. V6 (Labor Market Elasticity) approximates reabsorption as a single parameter, but historical technology transitions created entirely new industries and job categories that were unforeseeable ex ante. If AI generates a comparable wave of new human-complementary tasks, our displacement projections will overstate actual unemployment. This is the strongest counterargument to our central projections, and we take it seriously.
4. Feedback effects from displacement itself are modeled only as threshold triggers, not continuous dynamics. Mass unemployment reduces consumer spending, which reduces corporate revenue, which changes AI investment calculus. Political instability from displacement can produce regulatory responses that alter the adoption curve. These second-order effects are real but computationally intractable to model with current data. Our model captures them as discrete scenario switches, not smooth feedback loops.
5. Geographic, sectoral, and demographic granularity is limited. V1-V9 are calibrated primarily against US data. International displacement timelines differ due to labor cost structures, regulatory environments, and infrastructure readiness. Within the US, displacement will not be uniform across sectors, regions, or demographic groups. Our model produces national aggregates; local reality will vary substantially.
6. Historical precedent for forecast failures is sobering. The 2013 Frey & Osborne study estimated 47% of US jobs were at risk of automation -- a figure that became a media sensation but proved methodologically flawed (it measured task exposure, not economic viability or adoption probability). McKinsey's 2017 estimates required repeated revision. Our framework was designed to avoid these specific errors, but there are certainly errors we have not anticipated.
7. The model assumes rational economic actors. In practice, organizations make suboptimal decisions due to political dynamics, sunk-cost fallacies, vendor relationships, and executive ego. These irrational frictions may slow adoption beyond what V3 captures. Conversely, herd behavior and competitive panic may accelerate adoption beyond rational equilibrium. Both directions of error are possible.
8. Uncertainty ranges widen dramatically beyond 3-year horizons. Our 2026-2028 projections carry meaningful uncertainty bands (plus or minus 30-50%). Our 2029-2032 projections carry very wide uncertainty bands (plus or minus 100% or more on timing). Our 2033+ projections are scenario sketches, not forecasts. We present them because understanding the shape of possible futures is valuable even when precise timing is unknowable.
9. Black swan events are not modeled. A major geopolitical conflict disrupting semiconductor supply chains, a fundamental breakthrough in AI safety that produces voluntary deployment constraints, a global financial crisis that collapses AI investment, or an unexpected technical breakthrough that leapfrogs current architectures -- any of these would invalidate model parameters in ways that cannot be predicted from existing data. The model maps the space of futures consistent with current trends. It does not map all possible futures.