Global Automation Atlas

TEXT START: Automation affects the labour content of work differently across different contexts. Yet, most existing exposure measures assign fixed scores to tasks or occupations, limiting comparisons of automation exposure across countries.

URL SCAN: arXiv Abstract - "Global Automation Atlas"
FIRST LINE: [econ.GN > General Economics, submitted 16 May 2026]

I. THE DISSECTION

This paper is an exposure cartography exercise—it maps where automation lands on the global workforce, country by country, task by task. The framing is empirical and taxonomic: measure, classify, compare. On the surface, it reads as neutral data science. Underneath, it is a structural diagnosis dressed as methodology.

The paper's five findings are not news. They are the autopsy report:

1. Exposure is highly uneven and income-correlated (3.3% South Sudan vs. 61.6% China) — this is not a discovery. It is confirmation that automation pressure concentrates where productive work concentrates. Low-income countries have low exposure because they have low productive work to automate. This is a lullaby masquerading as data: "don't worry, the poor are less exposed." The truth is worse—they are less exposed because they are less integrated into the automated circuit. When they get integrated, the exposure will be catastrophic and sudden, not gradual.

2. Substitution skew is worse for low-income countries — This is the key signal. Low-income nations face substitution rather than augmentation. Augmentation means humans work alongside machines. Substitution means humans are redundant. The paper notes this asymmetry but treats it as a descriptive fact rather than what it is: a development trap with a timer attached. Low-income countries trying to climb the ladder will find that ladder is a treadmill moving in the opposite direction.

3. Less advanced automation dominates in low-income countries — The technology channel matters. "Simple" automation accounts for >50% of exposure in low-income settings. This means they face the blunt, job-destroying forms first. The assumption that countries will naturally climb the automation sophistication ladder is untested and likely false—AI is compressing that ladder into a single step that skips the intermediate stages.

4. AI is less prevalent in simpler automation but substitutes more in low-income settings, augments more in high-income settings — This is the central finding and it is devastating. AI does not arrive uniformly. It optimizes for the labor cost structure of the context. In low-income settings, where labor is cheap but not cheap enough, AI substitutes. In high-income settings, where labor is expensive, AI augments (makes humans more productive). This is not a neutral observation. It means AI will hollow out middle-income and lower-income workforces faster than it augmented high-income ones. The technology is regressive in its immediate impact.

5. Females disproportionately exposed to labor-substituting automation — Gendered labor market segmentation combined with substitution margins creates a compound disadvantage. This is structural, not incidental.

II. THE CORE FALLACY

The paper assumes that exposure can be managed, compared, and integrated into policy frameworks. The entire methodology is built on the premise that measuring automation exposure is the necessary first step toward managing it. This is the fallacy.

The Discontinuity Thesis identifies the core mechanism: AI severs the mass employment -> wage -> consumption circuit. This paper treats automation exposure as a variable that can be mapped, analyzed, and (implicitly) mitigated through informed policy. It does not ask the question that matters: what happens to the consumption circuit when exposure crosses a threshold in enough countries simultaneously?

The paper is structurally Keynesian in its assumptions—it implicitly assumes that labor will remain the primary distribution mechanism for purchasing power, and that understanding exposure gradients will allow interventions to smooth the transition. The DT rejects this. The lag defenses are real, but the direction is fixed. Mapping exposure more precisely does not change the structural outcome. It only makes the map of the disaster more legible.

The second fallacy is treating income-level as the primary determinant of exposure dynamics. The paper notes substantial within-group variation but treats it as noise. The actual mechanism is more specific: the competitive pressure on labor costs drives adoption. Countries that are integrated into global supply chains face immediate automation pressure regardless of income level, because the pressure comes from competitive cost optimization, not from internal development trajectory. China's 61.6% exposure is not because it is high-income—it is because it is the manufacturing center of a global system that is actively automating.

III. HIDDEN ASSUMPTIONS

1. That automation exposure is a metric that policy can act on. Implicitly assumes that national governments retain the institutional capacity to intervene meaningfully in automation adoption rates. Evidence from the past decade suggests this capacity is eroding rapidly as AI capabilities compress adoption timelines below political response cycles.

2. That the task-based approach captures the relevant unit of analysis. Tasks are what AI destroys when it achieves performance superiority. But the transition from task-automation to full-occupation-automation happens faster than the task framework can track. The paper builds a detailed map of a coastline that is actively dissolving.

3. That AI involvement is a separate dimension from substitution/augmentation. The paper treats these as orthogonal dimensions. They are not. Under the DT, AI involvement is the mechanism that determines which tasks survive long enough to be classified as "augmentable." The framing separates what is inseparable.

4. That the current distribution of exposure is stable. The paper treats 124 countries and 2.33 million labels as a snapshot. It is a photograph of a moving vehicle. The next photograph will show a different distribution, and the trajectory is not toward equilibrium.

IV. SOCIAL FUNCTION

This paper performs institutional legitimacy maintenance. It signals that economics is grappling seriously with automation, that measurement frameworks are being developed, that the discipline is responsive to structural change. This is transition management theater—creating the appearance of analytical engagement with a problem that the discipline's dominant frameworks cannot solve because the problem is systemic and the frameworks are equilibrium-based.

It is also policy seduction: the implication that better measurement leads to better policy responses gives policymakers an active role in the narrative. The uncomfortable truth is that the policy toolkit available to nation-states is structurally insufficient to address a system-level collapse in labor's economic function. UBI, transfers, industrial policy—these are lag defenses, not solutions. The paper gives them a technical veneer that makes them appear more effective than they are.

Secondary function: academic prestige signaling. Task-based, country-specific, 2.33 million data points—this is high-complexity empirical work that will generate citations and methodological spinoffs. The scholarly apparatus around it creates a sense of progress that is not matched by any genuine deceleration of the underlying process.

V. THE VERDICT

This paper is a precise, detailed, ultimately irrelevant map of a burning building.

It does excellent work within its own epistemic frame. The methodology is rigorous. The data coverage is impressive. The five findings are accurate as far as they go. But the frame itself is the problem: measuring exposure is not the same as diagnosing the disease, and treating the disease requires acknowledging that the patient is the entire economic order, not a set of vulnerable labor markets that can be smoothed with better data.

The paper's implicit optimism—that understanding automation exposure gradients will enable better policy—is the same optimism that produced Paris Agreement targets that have not been met, AI governance frameworks that have not slowed capability growth, and labor protection regulations that have not prevented the gig economy from hollowing out employment stability. Data enables action only when action is structurally possible. The DT's verdict on institutional response capacity is that it lags behind the pace of structural change by a margin that is growing, not shrinking.

Functional verdict: This paper is the academic version of rearranging deck chairs. Methodologically sophisticated, empirically valuable, structurally inconsequential. It will be widely cited, taught in graduate seminars, and cited in policy briefs that will not change the trajectory.

The question it should have asked: At what aggregate exposure threshold does the consumption circuit collapse, and is that threshold being approached faster than the measurement framework can track?

That question is not in the paper. And answering it would require abandoning the framework that makes the paper fundable and publishable.