As of February 2026, approximately 1% of US desk jobs have been fully replaced by AI systems. That number, by itself, suggests the disruption is a long way off. But displacement curves are not linear -- they are exponential at minimum, and potentially hyperbolic. The distance from 1% to 50% is shorter than it appears. How short depends on exactly three variables.
The Displacement Curve: Three Scenarios
We model desk job displacement under three distinct scenario frameworks. Each shares the same starting point -- roughly 1% full displacement in early 2026 -- but diverges sharply based on assumptions about AI capability growth, adoption friction, and the presence or absence of recursive self-improvement loops.
The conservative scenario assumes a fixed capability doubling period of 18-24 months, significant regulatory and institutional friction, and no meaningful recursive AI self-improvement. The base case assumes an accelerating doubling period (from ~18 months to ~10 months by 2028), moderate friction, and limited recursive effects. The recursive scenario assumes a collapsing doubling period driven by AI systems improving themselves, minimal adoption friction once competitive dynamics take hold, and a hyperbolic rather than exponential growth curve.
The chart above is the central thesis of this analysis. All three curves begin at the same point. Within four years they diverge by more than a decade. The reason for that divergence is not disagreement about whether AI will displace desk jobs -- on that point there is near-universal consensus. The disagreement is about the shape of the curve, and that shape is determined by exactly three variables.
Yearly Projection Data
The following table presents year-by-year displacement projections under each scenario. These figures represent the percentage of US desk jobs (approximately 60 million roles) that are fully replaced -- not augmented, not partially automated, but eliminated or consolidated to the point where the original headcount is no longer needed.
| YEAR | CONSERVATIVE | BASE CASE | RECURSIVE |
|---|---|---|---|
| 2024 | 0.2% | 0.2% | 0.2% |
| 2025 | 0.5% | 0.6% | 0.7% |
| 2026 | 1.0% | 1.0% | 1.0% |
| 2027 | 2.0% | 3.0% | 5.0% |
| 2028 | 3.5% | 8.0% | 17.0% |
| 2029 | 5.5% | 16.0% | 38.0% |
| 2030 | 8.0% | 28.0% | 58.0% |
| 2031 | 12.0% | 42.0% | 75.0% |
| 2032 | 16.0% | 55.0% | 86.0% |
| 2033 | 21.0% | 66.0% | 92.0% |
| 2034 | 27.0% | 74.0% | 95.0% |
| 2035 | 33.0% | 80.0% | 97.0% |
| 2036 | 38.0% | 85.0% | 98.0% |
| 2037 | 43.0% | 88.0% | 99.0% |
| 2038 | 48.0% | 91.0% | 99.0% |
| 2039 | 53.0% | 93.0% | 99.0% |
| 2040 | 57.0% | 94.0% | 99.0% |
| Highlighted values indicate the approximate year each scenario crosses the 50% threshold. Conservative: ~2039. Base Case: ~2031-2032. Recursive: ~2029-2030. Figures represent full displacement of roles, not task-level automation. | |||
The most striking feature of this table is the divergence after 2027. In 2026, all three scenarios are virtually identical -- about 1% displacement. By 2030, the gap between conservative (8%) and recursive (58%) is fifty percentage points. By 2034 the conservative model is where the recursive model was in 2029. That five-year lag represents approximately 16 million desk jobs -- the difference between a managed transition and a shock.
The Three Key Variables
Every displacement model, however sophisticated, ultimately reduces to three independent variables. These are not assumptions we chose for convenience -- they are the minimal set of parameters required to distinguish observed curve shapes. Every other factor (cost differentials, competitive pressure, workforce demographics) flows from these three.
Under the conservative scenario, AI self-improvement is real but bounded -- it speeds development by 20-30% but does not fundamentally alter the doubling period. Under the base case, the doubling period shrinks from ~18 months (2024) to ~10 months (2027) to ~7 months (2029). Under the recursive scenario, the doubling period collapses hyperbolically -- from 18 months to 5 months to 2 months -- approaching a mathematical singularity around 2029-2030.
Historical precedent supports the recursive model's assumption more than the conservative one. Digital photography displaced film photography in approximately 8 years despite enormous institutional resistance from Kodak and the broader film industry. Smartphones displaced feature phones in 6 years. The key insight: friction delays adoption curves but does not change their shape. Once the economic delta is large enough -- and at $5-15K/year for AI versus $80K/year per knowledge worker, it is -- friction becomes a timing factor, not a structural one.
The conservative scenario assumes this rate slows as AI encounters fundamental capability walls (similar to how Moore's Law eventually slowed for silicon). The base case assumes the current rate holds through 2030, then gradually decelerates as the easiest gains are captured. The recursive scenario assumes the base rate is sustained or accelerated by new architectural breakthroughs (reasoning models, multimodal integration, agentic frameworks) -- each new paradigm forming a new S-curve that maintains the overall exponential envelope.
These three variables are not independent of each other. Faster base capability growth makes recursive improvement more likely. Recursive improvement, once established, reduces adoption friction by making the economic case overwhelming. Lower friction accelerates deployment, which generates more training data and competitive pressure, which feeds back into base capability growth. This is why the recursive scenario is not simply "the optimistic one" -- it is the scenario in which the three variables reinforce each other.
The 50% Threshold: Historical Context
Why does 50% matter? Not because it represents the midpoint of a process, but because it represents a phase transition. When half of desk jobs are displaced, the event ceases to be an economic adjustment and becomes a structural transformation of society. Every prior transition that crossed the 50% threshold in a major employment category triggered cascading effects that were qualitatively different from the effects of incremental displacement.
EMPLOYMENT
90% → 2%
(1800-1900)
EMPLOYMENT
35% → 8%
(1960-2000)
DISPLACEMENT
1% → 50%
(2026-?)
The agricultural transition took nearly a century. The manufacturing transition took approximately forty years. Both were accompanied by massive social upheaval -- urbanization, labor movements, political realignment, new educational systems. Both eventually produced economies that were larger and more productive than what they replaced. But "eventually" is doing a lot of work in that sentence. The agricultural transition produced a civil war. The manufacturing transition produced the Great Depression and two world wars, not as direct results but as consequences of the same structural forces.
Agricultural revolution (1800-1900): Farm employment dropped from 90% to ~40% of the workforce over approximately 60 years. The 50% crossing point occurred around 1870. The transition drove urbanization, the Homestead Acts, railroad expansion, and ultimately the political realignment of the 1890s. Time from visible disruption to 50%: roughly 40 years.
Manufacturing displacement (1960-2000): Factory employment dropped from ~35% to ~8% of the workforce. The steepest decline occurred 1979-2000. Entire metropolitan economies collapsed within a decade. The "Rust Belt" phenomenon destroyed tax bases, school systems, and social infrastructure in communities that had been prosperous for generations. Time from peak to 50% decline: roughly 20 years.
AI desk job displacement (2026-?): Under the base case, the transition from 1% to 50% takes approximately 5-6 years. Under the recursive model, 3-4 years. This is not incrementally faster than prior transitions -- it is categorically different in velocity. No existing social safety net, retraining program, or institutional adaptation mechanism is designed to operate at this speed.
What Happens at 50%
The literature on economic phase transitions identifies several cascading effects that emerge specifically when displacement crosses the 50% threshold in a major employment category:
Consumer demand compression. Desk workers are also consumers. When half lose their positions within a compressed window, consumer spending drops even as corporate productivity rises. The GDP growth from AI efficiency is partially or fully offset by the demand destruction from mass displacement. This dynamic is not speculative -- it is precisely what occurred in manufacturing-dependent regions of the Midwest between 2000 and 2010.
Tax base collapse. Income tax, property tax, and sales tax revenue are all functions of employment. A 50% displacement of the highest-earning employment category (knowledge workers average $80K+ annually) represents a structural reduction in government revenue at every level -- federal, state, and municipal. This occurs precisely when demand for social services peaks.
Political instability. Both the agricultural and manufacturing transitions produced political realignment. The Populist movement of the 1890s, the New Deal coalition of the 1930s, and the post-industrial populism of the 2010s were all responses to employment displacement crossing critical thresholds. The AI transition is likely to produce a similar realignment, but compressed into a shorter timeframe, which historically correlates with more volatile political outcomes.
Institutional failure. Educational institutions, retirement systems, healthcare tied to employment, and housing markets all assume baseline employment levels. A 50% displacement of desk jobs would stress every one of these systems simultaneously. The 2008 financial crisis -- which involved approximately 8 million job losses over 24 months -- strained these systems to their limits. The AI transition, even under the conservative scenario, involves 30 million desk job losses over a decade. Under the recursive scenario, 30 million losses in 3-4 years.
The question is not whether the displacement happens -- the capability curves are too strong, the cost differential too large, the competitive dynamics too powerful. The question is whether we navigate the 50% threshold as a managed transition or an unmanaged shock. The difference between those outcomes is measured in years, and the clock is already running.
-- AI Labs analysis, Feb 2026
Which Scenario Is Most Likely?
As of February 2026, the empirical evidence is most consistent with the base case scenario, with early indicators suggesting the recursive scenario cannot be ruled out. Three data points are particularly informative:
METR benchmark acceleration. The overall capability doubling period measured by METR is ~7 months (2019-2026), but the post-2023 trend shows acceleration to ~4 months. This is consistent with the base case (accelerating but not yet hyperbolic) and weakly consistent with the early stages of the recursive scenario.
Corporate adoption speed. Block's 40% workforce reduction, announced February 2026, is the largest single AI-attributed layoff in history. Jack Dorsey's prediction that most companies will reach the same conclusion within 12 months is consistent with the low-friction assumptions of the base case and recursive scenarios, not the high-friction assumption of the conservative model.
AI self-authorship rates. Anthropic's confirmation that Claude writes 90% of its own code is the strongest single evidence point for the recursive scenario. If this capability generalizes across domains -- and there is no architectural reason it would not -- the recursive feedback loop is already operational, not theoretical.
Our current assessment: the base case is the most defensible central estimate, with the probability distribution skewed toward the recursive scenario rather than the conservative one. If forced to assign probabilities: Conservative 15%, Base Case 50%, Recursive 35%. The scenario that most people find intuitively comfortable -- the conservative one -- is, in our analysis, the least likely.
All projections use a composite model incorporating METR benchmark data (15 data points, 2019-2026), BLS occupational employment projections (2024-2034), Brynjolfsson/MIT ADP microdata on early-career displacement (2025), Goldman Sachs task-level automation estimates, and McKinsey capability-to-displacement conversion frameworks. The recursive scenario additionally incorporates Anthropic's self-authorship data and xAI co-founder Jimmy Ba's recursive loop timeline estimate. Full methodology and source data available in our methodology document and sources page. These are models, not predictions -- they represent structured estimates of plausible trajectories given current evidence.