Apollo chief economist says there's 'zero evidence' AI is killing jobs—it's creating them

URL SCAN: Apollo chief economist says there's 'zero evidence' AI is killing jobs—it's creating them

FIRST LINE: Major employers from Block to Amazon have attributed their shrinking workforces to new efficiencies in the tech revolution...

B. TEXT ANALYSIS PROTOCOL

1. THE DISSECTION

This article is a position paper dressed as journalism. It surfaces Sløk's thesis and then assembles a counter-narrative from sources (Forrester, Yale Budget Lab, Altman backtracking) designed to make the catastrophist position appear hysterical while burying the empirical signals that contradict it. The structure is a classic reassurance architecture: open with undeniable evidence (Amazon, Block, Klarna, Dimon) → pivot to expert who says the opposite → supply enough institutional cover (Altman "delighted to be wrong," Forrester's "AI washing" frame) to make the reassurance feel credible. The result is an article that appears balanced but functions as elite legitimization of cognitive dissonance.

2. THE CORE FALLACY

Sløk's fallacy: measuring displacement through current employment counts rather than output-per-worker ratios.

This is the central diagnostic error. Sløk cites ADP payroll data and nonfarm payrolls as evidence that AI isn't destroying jobs. But the article itself contains the refutation embedded in its own examples:

• Block: ~4,000 layoffs, same revenue trajectory, "smaller and flatter teams." This is output constant, headcount reduced. That is capital replacing labor. Not a productivity mystery—a displacement mechanism.
• Klarna: Let attrition do the work. 20% annual workforce reduction via natural turnover. No new hiring. Revenue and operations maintained with 20% fewer people in one year, 22% fewer by the following year. This is the precise mechanism by which mass employment collapses—not a layoff event, but a slow suffocation via attrition while output holds.
• Amazon: Jassy explicitly states they will need "fewer people doing some jobs" while maintaining operations. Sløk has no response to this.

The Jevons paradox framing—cheaper technology generates more demand for humans—is structurally invalid here because Jevons described physical energy efficiency gains (steam engines making coal more useful, prompting more coal use). AI is not a tool that enhances human productive capacity. It is a substitute for the cognitive labor that humans previously performed to generate value. The mechanism is categorically different. Sløk is applying a 19th-century industrial dynamics model to a 21st-century cognitive automation process.

3. HIDDEN ASSUMPTIONS

• Assumption 1: Employment counts capture displacement. They do not. A company can cut 20% of its workforce and grow output. That gap—headcount down, output stable—is the displacement signal, not its absence.
• Assumption 2: Current ADP payroll data is the relevant temporal window. But the DT mechanism operates on a lag structure. Automation displaces slowly, then all at once. Comparing 2026 payrolls to 2022 baselines misses the accumulation dynamic entirely. This is like noting clean air in 1975 as evidence that CFCs are harmless.
• Assumption 3: AI-related layoffs (55,000 in 2025 per NBER) represent the total displacement. This is the visible portion. The structural displacement from AI-driven productivity gains—where companies simply don't refill roles as people leave—doesn't appear as a layoff in layoff data. It appears as disappearing job reqs and headcount held flat while revenue grows. The article's own Klarna example demonstrates this: the company didn't announce mass layoffs. It simply stopped hiring. That displacement is invisible in layoff statistics.
• Assumption 4: Sløk's "hiring AI implementation experts" offsets the losses. This is a ** boutique labor market argument**. There are not enough "AI implementation experts" to absorb the clerical, analytical, creative, and coordination work performed by the hundreds of millions in middle-skill and high-skill knowledge work. The replacement ratio is not 1:1. It's not 10:1. It is indefinite: one AI system can perform the work previously requiring dozens of specialists, and the system improves without additional hiring.

4. SOCIAL FUNCTION

Classification: Institutional Confidence Maintenance / Investor-Reassurance Theater

The social function of this article is to provide institutional actors—LPs, pension funds, sovereign wealth vehicles, corporate boards—with verbal cover for the contradiction between what their portfolio companies are doing (cutting headcount, replacing humans, refusing to hire) and what they need to believe about the system's stability.

Sløk is the chief economist at Apollo Global Management, a firm managing ~$700 billion in assets. His function is not to predict economic disruption. His function is to provide the macroeconomic frame that keeps institutional investors confident in the returns of their existing capital structures. A man managing $700 billion in assets telling pension funds that AI won't kill jobs is not making a neutral economic argument. He is performing the reassurance function that his institutional clients require to remain committed.

The article then layers Altman's "delighted to be wrong" and Forrester's "AI washing" framing on top of this, which serves a displacement obfuscation function: if some layoffs are "AI washing" (using AI as a pretext for cuts they'd make anyway), then no layoff attributed to AI can be taken at face value. This allows the reader to dismiss the Amazon/Jassy signal as possibly pretextual, preserving the reassurance. This is epistemic laundering: the ambiguity argument is weaponized to maintain the null hypothesis that nothing structural is happening.

5. THE VERDICT

The article is a masterclass in lag-blind economism.

Sløk's "zero evidence" claim is only defensible if you define "evidence" as mass unemployment events visible in contemporaneous payroll data. But the Discontinuity Thesis does not predict mass unemployment events as the primary mechanism. It predicts productivity-output-rising-while-headcount-falls, which is exactly what the article's own examples demonstrate. The article tells you Block is doing more with fewer people. It tells you Klarna is shrinking its workforce at 20% annually without operational collapse. It tells you Amazon expects to need fewer corporate employees. Then it tells you Sløk says there's zero evidence of job loss.

The evidence Sløk dismisses is sitting in his own source material.

Under DT mechanics: the productivity gap is the displacement signal. The lag between productivity gains and employment collapse is the confirmation lag, not the absence of the phenomenon. Sløk is measuring the lag as proof the thing isn't happening. This is the same error as measuring atmospheric CO2 in 1985 and concluding the greenhouse effect was speculative.

The Jevons paradox does not apply to cognitive labor substitution. It applied to coal. AI is not making human cognitive work cheaper to use—it is making human cognitive work dispensable. The distinction is the entire thesis.

Structural Verdict: This article performs the social function of preventing institutional investors from pricing the discontinuity into their models. That is its purpose. Whether Sløk believes it is irrelevant—it is a structural communication from the asset management class that the current labor statistics are sufficient reassurance for capital commitment. The article is not wrong by accident. It is useful by design.