The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay

Sources#

• AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It
• Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence
• Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors

Summary#

Most of the vault's labor evidence is about headcount — who is employed (Canaries), which firms grow (Firm AI-Spend Intensity and Headcount Growth), which occupations post vacancies. This page holds the price side: whether AI exposure moves what a job pays. Two empirical instruments now bear on it, and read carelessly they disagree:

• Advertised pay rises faster in exposed occupations. Indeed Hiring Lab (Jack Kennedy, 2026-09-17) — median advertised annual salary in the most AI-exposed tercile is up ~46% since 2021 against ~25% in the least exposed, and a difference-in-differences puts the post-ChatGPT premium at +5.7%.
• Realized base pay barely moves. Brynjolfsson, Chandar & Chen (ADP payroll through June 2026), Fact 6: "little difference in compensation trends by age or exposure quintile" — adjustment runs through employment, not pay (detail on Task Saturation: Broad but Shallow AI Diffusion).

They are measuring different margins, and on the one they share they agree. The Indeed premium is a new-hire offer effect concentrated in senior roles, and most of it is composition — exposed occupations stopped posting entry-level jobs. Held at constant seniority, it falls to +2.4%, not significant.

Evidence note. empirical, but vendor data in the vendor's voice: Indeed's own salaried postings, analyzed and published by Indeed's research arm, as a blog post with no standard errors beyond significance stars. The exposure instrument (GSTI) is also Indeed's own. The sample is postings that advertise an annual salary, which skew heavily senior (≈37% senior in 2026 vs ≈14% across all US postings), so neither the levels nor the seniority shares generalize to the whole postings stream. The index is nominal — the less-exposed tercile's +25% since 2021 is roughly flat in real terms; the regressions absorb inflation with month fixed effects, the headline chart does not.

The instrument#

• Exposure: Indeed's GenAI Skill Transformation Index (GSTI), from its 2025 AI at Work report — GenAI models rated ~2,900 work skills on cognitive and physical demands and sorted each into minimal / assisted / hybrid / full transformation; an occupation's score is the share of its skills rated hybrid or full. That places it in Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated's LLM-as-rater family with Eloundou, not the telemetry family. Occupations are cut into equal-count terciles; top vs bottom third is the headline. High: software development, IT systems & support, data & analytics, marketing, banking & finance. Low: nursing, personal care, food service, cleaning, manufacturing. The same instrument underlies Gallacher's July 2026 postings-count analysis on Firm AI-Spend Intensity and Headcount Growth.
• Outcome: median advertised annual salary per posting, indexed to each occupation's 2021 average, with occupation shares fixed at 2021 — so the headline already removes cross-occupation mix (the Software-Development-vs-Driving shift), but not the within-occupation seniority mix.
• Seniority: Indeed's extracted entry/mid/senior posting attribute; ≈99.5% of salaried postings carry exactly one level tag every year.

Three specifications, ordered by how much seniority shift they remove#

Model	Pay measure	Fixed effects	Seniority mix	Premium (exposure × post-ChatGPT)
1	Median	Occupation + month	Not held	+5.7% (p<.05)
2	Mean of log	Job title + month	Held in part (~half)	+4.7% (p<.01)
3	Median	Occupation × level + month	Held	+2.4% (n.s.)

SEs clustered by occupation. The author stresses these are alternative cuts, not a nested sequence. The footnote (image-only in the original; verified against the image) supplies the like-for-like comparisons: switching Model 1 to mean-of-log alone gives +6.3%, and Model 1 run on Model 3's seniority-level cells gives +6.2% — so the narrowing attributable to holding seniority is 6.2% → 2.4%, which is most of the effect.

Kennedy's defence of the larger numbers is fair and should travel with the small one: if AI reshaping tasks is why exposed occupations tilted senior, Model 3 "may tend to over-correct, removing some of what AI is doing rather than a distortion in the data." Composition is not a nuisance here; it may be the treatment. What it is not is a raise — nobody doing the same job at the same level is being offered significantly more.

The composition effect is the entry-level finding#

The sentence in the post that matters most for this vault is in its caveats: in the most-exposed occupations, the entry-level share of salaried postings fell from 29% to 10% between 2021 and 2026, while the senior share rose from 22% to 47%. In the least exposed, the tilt was about a third as large.

That is a vacancy-flow reading of the same event Canaries reads in payroll — exposed occupations cutting the bottom rung through hiring, not separations — on an independent data source, extended to 2026, and at a magnitude (entry share cut by two-thirds) far larger than the ~20-point employment divergence. It also agrees with Gallacher's finding that 71% of the 2025–26 software-postings rebound was senior roles (Firm AI-Spend Intensity and Headcount Growth). Three limits: it is a share, not a count, of a total that fell sharply then rebounded; it covers only salary-advertising postings; and the window starts in 2021, so part of the tilt may predate ChatGPT (the post does not split it). What it adds to The Tragedy of the Cognitive Commons is persistence on a second instrument — nothing about whether the remaining entry roles still teach.

The seniority gradient is mostly pre-ChatGPT#

The key points say the premium "widens with seniority" — senior +17 points (145 vs 128 on the 2021 index), mid ≈+12 (132 vs 120), entry ≈+2 (124 vs 122). The post then separates two questions and the second one deflates the first: measured from a 2022 baseline, the post-ChatGPT gap is roughly 5 points at entry, 6 at mid, 7 at senior — "far more even across levels." So most of the cumulative seniority gradient was already in place by 2022, and the regression terms behind the split are "mostly not statistically significant"; the author treats it as suggestive. Anyone quoting "the premium rises with seniority" as an AI effect is quoting the 2021 baseline.

Timing: the more- and less-exposed lines tracked each other for about a year after ChatGPT; the gap opened around 2024 — coinciding, the author notes, with the rebound in exposed-occupation postings. The less-exposed index peaked near 128 around mid-2025 and has drifted down since (chart reading, approximate).

Reconciling with realized pay (Canaries Fact 6)#

Match the margins and the two sources say one thing:

Margin	Canaries (ADP, real base pay, through Jun 2026)	Indeed (advertised offers, nominal, through mid-2026)
Young / entry new hires	Starting pay: no relationship with exposure	Entry gap ≈2 points cumulative; entry panel's less-exposed line at or above for most of the period
Older / senior new hires	Starting pay rose modestly faster in exposed jobs	Senior gap largest (+17 cumulative, ≈7 post-2022)
Incumbents (job-stayers)	Real base pay grew slightly slower in exposed jobs	Not observed
Pooled, composition-adjusted	Little divergence; composition flagged as a distortion risk	+2.4% n.s. with seniority held

So: a modest senior-hire premium, no entry-level premium, nothing for incumbents, and a headline number inflated by who is being hired. Canaries anticipated exactly this distortion ("if hiring of young exposed workers contracts, the remaining pool skews toward higher-tenure, higher-paid incumbents"); Indeed's Model 3 is the postings-side measurement of it. Neither sees bonuses or equity, which Canaries notes are largest in the most-exposed, high-income occupations.

What both instruments miss: the graduates who did not get the job. Both condition on being in, or posting for, the exposed occupation. Orr, Tucker & Warren (US Census Bureau, September 2026) (empirical, PSEO×LEHD) assign exposure by college major instead. Graduates of the most-exposed decile of majors lose 13% of initial full-quarter earnings, and about half of that is sorting into lower-paying sectors (retail, food service). So the flat entry row above does not mean the entry-level wage margin is untouched. The price of an exposed entry job barely moved, and fewer graduates got one. Detail on AI-Exposed College Majors at Labor-Market Entry.

Under Autor & Thompson (the model Task Saturation: Broad but Shallow AI Diffusion uses for the crossover signal): automating an occupation's inexpert tasks lowers employment and raises the scarcity value of remaining expertise; automating expert tasks depresses wages. Falling entry share plus a senior-offer premium is the inexpert-automation signature on the offer margin — directionally the same reading as Codified vs Tacit Knowledge Exposure's substitution-for-the-young / complement-for-the-experienced gradient. It is weak evidence: the seniority split is not significant, and the post-2022 gap is nearly flat across levels.

Kennedy's conclusion outruns the data. "AI has been acting more as a complement to skilled workers than a replacement" and "AI skills appear to command a premium" are interpretations; the post does not observe AI skills in postings, and the only significant estimates are the ones that do not hold seniority fixed. Its employer advice (update pay benchmarks, advertise AI skills) is a recommendation from the platform that sells job ads.

Connections#

• Seniority-Biased AI Adoption: The Junior Share at Adopting Firms — the same senior tilt measured in headcount at AI-adopting firms rather than in postings: junior share −1.9pp, senior employment +6.7%, across 41 countries. That paper cannot see wages, and this page is the vault's only offer-side reading of them
• Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — GSTI is one more LLM-rated instrument, and this is the first realized wage outcome scored against any exposure instrument in the vault. It extends that page's "the sign depends on the instrument and the window" to "and on the composition control": positive and significant without seniority held, insignificant with it
• The Tragedy of the Cognitive Commons — the entry-level posting share falling 29% → 10% in exposed occupations is the regeneration slot disappearing at the vacancy stage, on a non-ADP instrument through 2026
• Task Saturation: Broad but Shallow AI Diffusion — the Autor–Thompson crossover signal (relative pay of exposed occupations by career stage) gets its second reading here, on the offer margin, and it agrees with the first
• Returns to Expertise in Agentic Coding — the senior-offer premium is the expertise premium priced at hire, though mostly as a pre-2022 gradient rather than a post-ChatGPT one
• Codified vs Tacit Knowledge Exposure — same age/seniority shape on a different outcome (offered pay rather than headcount), with the same caution that the gradient's AI-attributable part is small
• Firm AI-Spend Intensity and Headcount Growth — home of the Indeed postings-count series and its contradiction with Ramp on entry-level composition; this page adds that the entry share of exposed salaried postings kept falling through 2026
• Stanford Digital Economy Lab — Canaries' instrument caveats, including base-pay-only compensation
• AI-Exposed College Majors at Labor-Market Entry — the entry-level earnings loss that offer and payroll data cannot see, because it happens through which sector graduates land in: −13% initial earnings for the most-exposed decile of majors, about half from sorting

Open Questions#

• Is the post-ChatGPT senior-offer premium real once seniority is held and the sample is powered for it? Model 3's +2.4% is not significant and the level split's terms mostly are not. Falsifiable: a matched-title, matched-level panel of advertised or realized new-hire pay by exposure, with bonus/equity included, over a window ending 2027 or later.
• How much of the 29% → 10% entry-share collapse in exposed occupations predates November 2022? The post reports only 2021 and 2026 endpoints; a 2022-baseline split would say whether the tilt is a post-ChatGPT event or a continuation of the post-COVID tech correction that Canaries' pre-trend decomposition also has to handle. Related (2026-10-01), not an answer: on graduates rather than postings, AI-Exposed College Majors at Labor-Market Entry finds the two margins split. Initial employment for the most-exposed majors is flat before 2020 and breaks right after November 2022. Initial earnings were already below their pre-pandemic level relative to less-exposed majors by 2022. So on that instrument the hiring break is post-ChatGPT and part of the pay decline is not.

Sources#

• AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It — Jack Kennedy, AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It (Indeed Hiring Lab, 2026-09-17; empirical, the platform's own salaried-postings data, blog post). Key points, §"AI-exposed roles have pulled ahead on pay" (46% / 25%, 5.7% / 4.7% / 2.4%), §"How do the trends differ by seniority?" (17 / 12 / 2 cumulative; 5 / 6 / 7 post-2022), §"What might be driving this?" (29% → 10% entry share, 22% → 47% senior share), Appendix and Methodology (GSTI terciles, salaried-postings restriction, 37% vs 14% senior share). Parse notes: the three figures are images, transcribed at ingest; chart values other than printed endpoints are approximate and are not quoted here except as flagged. The regression table and its footnote exist only in the image and were re-read from it (transcription matches). Source-internal inconsistency, unreconciled: the footnote says the senior share of most-exposed postings rose +16.3pp, of which +10.5pp occurred between job titles and +11.4pp within them — the parts sum to 21.9, not 16.3 — while the prose gives the senior share rising 22% → 47% (+25pp). The windows or weighting presumably differ (the footnote may be post-2022 or occupation-share-fixed), but the post does not say; do not cite the footnote decomposition
• Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence — Brynjolfsson, Chandar & Chen, revised August 2026 (empirical, ADP payroll through June 2026). Cited only for §2.6 Fact 6 (compensation margin, job-stayer and new-hire splits, the composition caveat, base-pay-only limitation) and footnote 22 (software developers' young-worker pay as an idiosyncratic exception). Full treatment on Task Saturation: Broad but Shallow AI Diffusion and The Tragedy of the Cognitive Commons
• Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors — Orr, Tucker & Warren, Graduating into Disruption (US Census Bureau CES-WP-26-56, September 2026, empirical). Cited for the major-level entry-earnings result (§1, §4.1.2–4.1.3) and the employment/earnings pre-trend split (§4.1); full treatment on AI-Exposed College Majors at Labor-Market Entry