A data-journalism investigation · April 2026
In search of
the AI throne.
The largest U.S. technology companies have committed trillions of dollars to AI infrastructure since the deep-learning era. Whether that bill ever returns its cost in profit is now a real debate — influential financial analyses argue the revenue-to-capex math may never close. A widely circulating thesis gives a different reading of what is being purchased: not a product line whose margins must eventually arrive, but a permanent capital advantage that dismantles high-paid cognitive work first. This piece walks the U.S. labor record from the deep-learning era to 2023 to see whether that throne is visible in the public data — and where it isn't.
We don't try to settle the thesis. We follow it through three questions: where AI is positioned in the labor market, whether that positioning has translated into displacement, and where the public record falls silent. You leave with a watchlist, not a verdict.
Act 1 · Where AI is positioned
The threat model has a measurable basis.
The Felten–Raj–Seamans AI Occupational Exposure index scores 774 detailed U.S. occupations on how much of their task content current AI systems can perform. Cross-referenced with the BLS wage panel for 2018 — the last full year before the SOC 2010 → 2018 vintage break, leaving 749 occupations on both sides of the join — one fact stands out: AI is positioned over high-wage cognitive work. The correlation between AIOE and log mean wage is +0.58; the correlation with employment count is essentially zero (+0.08). That is exactly the substitution surface the displacement thesis would predict.
Figure 08
AI exposure × wages × employment, 2018
Each point is one detailed occupation in 2018. The cloud bends up and to the right: AI exposure rises with wages (r = +0.58), and barely at all with headcount. This is the substitution surface the throne thesis would predict — high-paid cognitive work in the line of fire. It is position, not event; whether the position has produced an event is what Act 2 asks of the post-2010 employment record.
Exposure is not uniform across the economy. The AIOE distribution has a heavy right-tail concentration in cognitive-occupation families — legal, business and financial operations, computer and mathematical, education and training — while production, transportation, construction, and personal-care families sit on the low side. A "lawyer" is one of the most-exposed occupations in the country and a "courier" is one of the least — and within a single occupation family, the AIOE spread can rival the gap between families.
Figure 04
How AI exposure is distributed across U.S. occupations
The AIOE distribution is centered near zero (median = −0.05), but the visible feature is a sharp spike on the right tail at AIOE ≈ +1.3 — lawyers, financial analysts, software engineers, market researchers. That spike is where the throne thesis lives. The roughly three-quarters of detailed occupations that fall outside the top quartile aren't the story the thesis is telling. The investigation that follows works on that long tail, not on the bulk.
Figure 05
Occupation families differ less in median exposure than they do inside themselves.
Inside a single SOC family the AIOE spread can rival the gap between families — Healthcare Practitioners alone covers nearly the full AIOE scale, and Office & Admin Support holds occupations on both ends. Family medians smooth those differences out and lose the cohort the thesis is actually about. This is the methodological reason the rest of the piece works on AIOE quartiles of detailed occupations — a family-level read would average the throne away before the test ever ran.
Act 2 · Through 2023
Displacement isn't visible at scale.
The thesis predicts that high-AIOE work shrinks first. The CES national series, indexed to January 2010, says the opposite. Professional & business services (high AIOE) is up roughly +35%; education & health services is up +40%; manufacturing (low AIOE) is up about +12%. The pandemic registers as a single-period shock, not a regime change. Through 2023 the threat-model surface and the trend-line are not the same thing.
Figure 01
Indexed sector employment, Jan 2010 = 100
If the throne thesis were already showing up at sector scale, the high-AIOE band — Professional & business services — should be the one bending down or stalling by 2023. PBS does plateau after its 2022 peak (index ≈ 138), drifting down to +35% by 2026 — a faint stall, not a collapse. The actual top performer is Education & health (+40%), and the only sector visibly weakening into the late panel is Manufacturing (+12%) — low-AIOE, the wrong sector for the thesis. The pandemic registers as a single shock, not a regime change. At this scale of resolution, no clean displacement event is visible.
Year-on-year change makes the same point in a different register. After the COVID spike washes out, services drift positive into 2026 while manufacturing settles into a sustained negative YoY band from 2024 onward. The sector that weakens is the low-AIOE one, not the high-AIOE one — the opposite of what a sector-level AI displacement story would predict.
Figure 02
Sector momentum, year-over-year
A sector-scale AI displacement shock between 2021 and 2023 should appear as a sustained negative band in the high-AIOE panel — Professional & business services. It doesn't. The negative band that is visible after 2023 sits in Manufacturing, the low-AIOE sector, where YoY growth has been below zero from early 2024 through the end of the panel. Education & health drifts positive throughout. The yearly cadence does not register the event the thesis predicts at sector scale — which is why Act 3 abandons the sector view and zooms to the occupation level.
Figure 03
Structural restructuring is decades old.
Restructuring is the baseline against which any AI signal would have to be detected. Across the deep-learning era window (2010–2026), Education & health gained +2.2 pp of total-nonfarm share (+15% relative), Professional & business services gained +1.3 pp (+10% relative, with a 2022 peak it has not regained), and Manufacturing lost 0.4 pp (−8% relative). The pattern was already in motion when this window opened — AI didn't start the reshuffle. The question Act 3 carries forward is whether anything inside the high-AIOE cohort separates an AI signal from that long-running trend.
Act 3 · Probing the thesis
Move the threshold. Watch the band.
If the thesis has a visible signature, it should appear when we isolate the most-exposed occupations and ask whether their wages are compressing. The OEWS panel, anchor years 2012 through 2023, tells a thinner story than the headlines suggest: median wages drift up, and inside the high-AIOE cohort — the one the throne thesis is actually about — wage dispersion barely moves at all. The 2018 → 2021 break in the figure marks the SOC 2010 → SOC 2018 vintage change, not a labor-market event.
Figure 10 · cohort map, 2018
Before we test the throne, here are the four cohorts.
The thesis test in the next chart sorts occupations into AI-exposure quartiles and asks whether their wage dispersion has compressed since 2012. Before that, the reader needs to see what each quartile is. Sweep Q1 → Q4: the wage distribution shifts right (median goes from $40k to $82k), and the SOC mix moves from production, construction, and transportation work in Q1 toward Education & Library, Business & Financial, Office & Administrative, Life/Physical/Social Sciences, Management, and Computer & Math in Q4. The throne thesis usually invokes Computer & Math, Legal, and Finance work specifically — but the AIOE Q4 quartile is broader than that, and that's the cohort the rest of the piece tests.
Why 2018? Felten et al. publish AIOE on SOC 2010, and 2018 is the last clean SOC 2010 OEWS panel before the SOC 2018 vintage flip. The merge is unambiguous in 2018, and the same cohorts persist as the longitudinal sample tested in Figure 09.
Figure 06
Wages across the deep-learning era
Each box is the cross-occupation wage distribution in May of that year. Two readings sit together: medians drift up steadily — a roughly 34% nominal climb over the window, in the neighborhood of CPI — and the interquartile body of the distribution holds its shape. The upper whisker does extend: the highest-paid occupations pull further from the median in nominal terms across the panel. But the aggregate cannot tell us whether that upper-tail movement is AI-driven or just the post-COVID wage cycle, and at this resolution it cannot test the thesis at all. The 2018 → 2021 break marks the SOC 2010 → SOC 2018 vintage flip, not a labor-market event. The aggregate, then, is too coarse for the thesis test — which is why we now split it.
A per-occupation view of nominal wage growth from 2012 to 2018 — the last clean SOC 2010 window — sits almost entirely to the right of zero. The 1,344 matched SOC 2010 occupations cluster around a median of +12.8% (p10 = +7.1%, p90 = +20.0%), with only a thin negative tail. The bulk of the labor market grew in nominal terms over those six years. Real (CPI-deflated) wages are an open question we flag explicitly.
Figure 07
Wage growth, occupation by occupation
If the displacement story were already underway by 2018, more of this distribution should be sitting to the left of zero. It isn't. The 1,344 SOC 2010 occupations cluster around a +12.8% nominal-growth median (p10 = +7.1%, p90 = +20.0%), with only a thin negative tail. This aggregate view does not split occupations by AIOE — that's the role of Figure 09. The shape here just establishes the baseline: nominal wages broadly grew across the labor market through 2018, and any AI-specific compression has to be visible against that baseline. CPI-deflated real wages are a separate question we flag explicitly.
The thesis we kept hearing in every conversation about this dataset has a sharper form: that the people who own the AI are using it to compress the people who used to be paid the most for thinking. That is a falsifiable claim. Sort occupations into quartiles by AI exposure, hold those quartiles fixed, and ask one question: from 2012 to 2023, did the cross-occupation p90/p50 wage ratio inside each quartile narrow? If the throne is being assembled, the high-AIOE row (Q4) should land deepest in negative territory. The low-AIOE row should sit closest to zero.
Figure 09 · testing the throne
If the throne were being built, the red row would have moved left
One row per AIOE quartile. Each line is the change from 2012 to 2023 in the cross-occupation p90/p50 wage ratio — bars to the left of zero mean the high earners and the median earners moved closer together inside that quartile. The throne thesis predicts the red row (Q4, the most-exposed cognitive cohort) lands deepest in the pink zone. It sits on the zero line. The two rows that did compress are low-AIOE work — the opposite of the prediction. This is the visible answer in the public panel; what the panel cannot see is what we hand to the coda as a watchlist.
Two readings are honest at once. First: through the public 2023 panel, the visible signature of the throne thesis is absent. The high-AIOE row didn't move (+0.002). The two rows that did compress — low-AIOE Q1 (−0.061) and Q2 (−0.058) — are the opposite of what a top-down AI substitution story would predict. Second: this is a coarse instrument. The OEWS p90 is top-coded near $208,000, broad LLM deployment did not arrive until late 2022, and the 2023 panel is restricted to occupations whose SOC 2010 code survived into SOC 2018. The chart can rule the thesis out at the level of the public panel; it cannot rule it out for the executive tier above the top-code or for the year that has not yet been released.
Coda · What you should watch
You leave with a watchlist, not a verdict.
Our window closes in 2023, a year before broad LLM deployment shows up in the federal panels. If the substitution-of-elites thesis is going to leave a visible signature, it should arrive in the next OEWS release and the next JOLTS series for high-AIOE occupations. Two columns — what we've found, and what we're watching for — sit side by side because they have to.
What the data say
AI is positioned over high-wage cognitive work (correlation +0.58 between AIOE and log wages, Fig 08). Through 2023, that positioning has not produced the wage compression the displacement thesis predicts inside the high-AIOE quartile (Fig 09).
Through the 2010–2026 window, Education & health gained +2.2 pp of total-nonfarm share and Professional & business services gained +1.3 pp, while Manufacturing lost 0.4 pp. AI is overlaid on a reshuffle already in motion. COVID was a spike, not a regime change.
In the high-exposure quartile, the cross-occupation p90/p50 wage ratio is essentially identical in 2023 and 2012. The throne thesis has no signature in the public panel through 2023.
What to watch next
Openings and separations in Computer & Math and Legal — the strongest single test the thesis can survive in public data.
The thesis is fundamentally about capital share. BLS occupation data can't see it; BEA factor-share series and Piketty/Saez/Zucman top-income panels can.
Our panel ends 2023 — a year before broad LLM deployment. The next OEWS release is the first window into post-LLM wage dynamics.