Adoption is the bottleneck, not capability
Karpathy's treemap scores what AI could automate. Anthropic's data shows what it actually is automating. The gap between the two is massive — and it means most AI disruption hasn't happened yet.
Theoretical Feasibility vs Observed Coverage
Each bar shows two values: the faded bar is how much AI could theoretically do (from O*NET task analysis). The solid bar is how much it's actually doing (from real Claude usage data). The gap between them is unrealized disruption potential.
Actual coverage (what's happening now)
Theoretical feasibility (what's possible)
Computer & Math3x gap — most adoption, still only 1/3 realized
Business & Finance3.4x gap — $3.7T in exposed wages
Management4x gap — high exposure, slow uptake
Office & Admin3.6x gap — automation candidates
Legal5x gap — regulatory friction slows adoption
Architecture & Engineering5.7x gap — physical deliverables protect
Arts & Media7x gap — largest unrealized potential
Why the gap matters
97%
of Claude usage is on feasible tasks
When people do use AI, they use it for things it can actually do. The problem isn't that AI fails at tasks — it's that most organizations haven't adopted it yet.
Conclusion: the disruption wave hasn't crested. It's still building.
Highest-coverage occupations today
Computer Programmers75% coverage
Data Entry Keyers67% coverage
Customer Service (API)High coverage
Med. TranscriptionistsKarpathy: 10/10
Anthropic
Karpathy
Common thread: pure digital I/O. No physical component. Measurable output.
- AI disruption predictions based on capability alone are premature — actual adoption is 3-7x behind what's technically feasible
- The occupations with highest actual coverage (programmers, data entry, customer service) share a trait: their output is digital and measurable
- This gap will close. When it does, the impact on the categories above will be significant — but we don't know how fast
- 30% of workers have zero AI coverage today (cooks, mechanics, bartenders) — physical work remains a hard boundary
AI targets the well-paid, not the vulnerable
The standard automation narrative says AI replaces low-skill workers first. The data says the opposite: the most exposed workers are educated, high-earning, and older. This is "measurability-biased technical change."
+47%
Income Premium
Highly exposed workers earn 47% more on average than unexposed workers
This isn't a low-wage problem. AI targets the execution premium in knowledge work.
17.4%
Hold Graduate Degrees
vs 4.5% in the unexposed group — nearly 4x the rate
Education doesn't protect you. It correlates with doing measurable cognitive work.
56%
AI Skills Wage Premium
Jobs requiring AI skills pay 56% more, up from 25% the prior year (PwC 2025)
The premium for wielding AI is rising as fast as the premium for doing what AI does is threatened.
$3.7T
Wages in High-Exposure Jobs
Annual wages in jobs scoring 7+ on Karpathy's 0-10 scale
That's the economic stake. Trillions in wages tied to tasks AI can already do.
The Exposed Worker Profile
"The fault line is no longer how educated you are. It's whether your output can be measured. If it can, it will be industrialized."
— Christian Catalini (@ccatalini)
The Measurability Thesis
Catalini argues this is "measurability-biased technical change." What's exposed isn't "strategy" or "research" — it's the measurable execution bundled inside those roles: the drafts, the models, the analyses. The economy built a wage premium around that bundle. AI unbundles it. What survives is the non-measurable core: defining intent, navigating genuine uncertainty, verifying agentic work, and absorbing liability.
- AI exposure inverts the traditional automation narrative — it targets the educated and well-paid, not the vulnerable
- The common denominator is measurability of output, not skill level — if your deliverable can be specified and verified digitally, it's exposed
- Two-speed economy emerging: the premium for doing what AI does is eroding, while the premium for wielding AI doubles year-over-year (25% → 56%)
- Your defense isn't more education — it's shifting from execution (measurable) to judgment (non-measurable)
No mass unemployment — but the on-ramp is narrowing
The headline: experienced workers are fine. The buried lede: young workers entering AI-exposed occupations are not. The labor market isn't collapsing — it's selectively closing doors for new entrants.
The Age Inversion
Same occupations, same AI exposure, opposite outcomes depending on age. AI augments incumbents who know what to ask for. It replaces the grunt work that used to train newcomers.
-13%
Ages 22-25
Employment in AI-exposed jobs
The entry ramp is shrinking
vs
+6-13%
Ages 30+
Employment in same jobs
AI amplifies experienced workers
Stanford/MIT 2025
Post-ChatGPT era
Timeline — how we got here
Nov 2022
ChatGPT launches. No immediate labor market signal.
2023-2024
AI adoption doubles (3.7% → 9.7% of U.S. firms). Young worker hiring quietly slows in exposed occupations.
Aug 2025
First hard evidence: Stanford study finds 13% relative employment decline for early-career workers in most AI-exposed jobs.
Jan 2026
Anthropic Economic Index: augmentation (52%) overtakes automation (45%). Workers are using AI to do more, not fewer of them doing it — yet.
Mar 2026
Anthropic labor study confirms: no systematic unemployment divergence. But BLS projects weaker growth for exposed occupations through 2034. The slope is gradual but clear.
The Growth Forecast
-0.6pp
Per 10pp Coverage Increase
For every 10 percentage-point increase in observed AI coverage, BLS employment growth projections decrease by 0.6 percentage points through 2034.
70% of low-exposure jobsRising early-career employment
<50% of high-exposure jobsRising early-career employment
This is a slope, not a cliff. But the direction is unambiguous.
- There is no AI-driven mass unemployment in 2026 — the "robots took our jobs" narrative is not supported by current data
- But there IS a measurable hiring slowdown for young workers (22-25) entering AI-exposed occupations (-13%), while experienced workers thrive (+6-13%)
- The mechanism: AI replaces entry-level tasks that used to be the training ground. Seniors leverage AI; juniors compete with it
- This is a leading indicator, not the full impact. Adoption is still early (see Tab 1). As coverage increases, the -0.6pp/10pp relationship suggests growing pressure
- If you're early-career in a high-exposure field: learn to steer AI, not compete with it. The verification/judgment layer is what's hiring
The data contradicts itself — and that's the point
Five paradoxes that define this moment. Each one shows why simplistic "AI will/won't take jobs" takes are wrong. The reality is a set of unresolved tensions that will play out over the next decade.
1. The Capability Gap
60-94% feasible
→
<10% displacement
AI can theoretically automate the majority of knowledge work. Actual job displacement is near zero.
Verdict: adoption lag buys time, but the ceiling is high. Don't mistake slow adoption for low risk.
2. The Productivity Paradox
+30% task gains
→
~0% GDP impact
Individual workers report massive productivity boosts. Goldman Sachs says the economy-wide signal is "basically zero."
Verdict: gains are real but localized. Economy-wide impact requires broader adoption (see Tab 1).
3. The Age Inversion
Seniors: +6-13%
→
Juniors: -13%
Experienced workers in AI-exposed jobs are thriving. Entry-level workers in the same jobs are declining.
Verdict: AI augments people who know what to verify. It replaces the work that used to teach you how.
4. The Augmentation Flip
52% augmentation
→
45% automation
In Jan 2026, augmentation overtook automation as the primary AI use pattern. Workers are using AI to do more, not to do less.
Verdict: today's augmentation is tomorrow's automation. The pattern flipped once already and could flip again.
5. The Growth Contradiction
Most automatable roles
→
Still growing (+38%)
Even jobs scored highest for AI exposure are showing net employment growth. Low-exposure jobs grew faster (+65%), but nobody's shrinking yet.
Verdict: net growth doesn't mean no disruption. The mix within each role is changing even if headcount isn't.
"What survives is the non-measurable core: defining intent, navigating genuine uncertainty, verifying agentic work, and absorbing liability. The question for every knowledge worker is simple: strip away the execution layer, and what's left? That's your moat. Or your problem."
— Christian Catalini (@ccatalini)
- Short term (2026): No mass unemployment. But entry-level hiring is already tightening in exposed fields. Act now if you're early-career.
- Medium term (2027-2030): The 3-7x adoption gap will narrow. As it does, the -0.6pp/10pp growth relationship predicts accelerating pressure on exposed occupations.
- The individual question: What percentage of your job is measurable execution (drafts, analysis, models) vs. non-measurable judgment (intent, verification, liability)? That ratio is your exposure.
- The structural shift: From doing → steering. From executing → verifying. From producing answers → underwriting their consequences. This is the transition already underway.
- Historical caveat: Offshoring predictions identified "a quarter of US jobs as vulnerable" — they proved largely wrong within a decade. But this time, we have real usage telemetry, and the gap is closing.