---
title: "Same Storm, Different Boats: Generative AI and the Age Gradient in Hiring"
aliases:
  - "Same Storm, Different Boats"
  - "Lodefalk et al 2026"
  - "Sweden Age Gradient"
tags:
  - sweden
  - eu
  - causal
  - age-gradient
  - hiring
  - daioe-index
  - working-paper
source: "/research/2026-03-orebro-sweden-age-gradient.pdf"
extract: "/research/extracts/2026-03-orebro-sweden-age-gradient.md"
authors: "Magnus Lodefalk; Lydia Löthman; Michael Koch; Erik Engberg"
publisher: "Örebro University School of Business"
date: "2026-03-16"
---

# Same Storm, Different Boats: Generative AI and the Age Gradient in Hiring

> [!info] Source
> PDF (human): [2026-03-orebro-sweden-age-gradient.pdf](/research/2026-03-orebro-sweden-age-gradient.pdf) · Raw extract (machine): [2026-03-orebro-sweden-age-gradient.md](/research/extracts/2026-03-orebro-sweden-age-gradient.md)
> Örebro Working Paper 2/2026 (Economics) — 16 March 2026 — Empirical study using Swedish population-register data, 58 pages
> **JEL:** J23, J24, O33 · **ISSN:** 1403-0586
> **Keywords:** Generative AI, job postings, labour demand, employment composition, monetary policy

## TL;DR

Uses Sweden's natural experiment — the **Riksbank rate hike (April 2022) preceded the ChatGPT launch (Nov 2022) by seven months** — to disentangle monetary tightening from AI as the cause of the post-2022 hiring slowdown. Finding: **the aggregate posting decline aligns with monetary policy, not AI**. But within employers, **22–25 year olds in high-AI-exposure occupations saw employment fall 5.5% by early 2025** relative to less-exposed occupations within the same employers, while **employees aged 50+ rose 1.3%**. Generative AI reshapes hiring **composition**, not aggregate demand, with the adjustment burden falling disproportionately on entry-level workers. **First population-register evidence of the "canaries in the coal mine" age gradient outside the US.**

## Key findings

### Aggregate posting decline is monetary, not AI

- Posting decline began with the Riksbank's first rate hike in April 2022, **seven months before** ChatGPT.
- Difference-in-differences: Riksbank rate-hike interaction precisely estimated (β̂₁ = −0.127, p < 0.01). ChatGPT-period AI-exposure interaction small and not significant (β̂₂ = −0.062, p = 0.11).
- AI exposure uncorrelated with monetary policy sensitivity (r = 0.04, p = 0.51) — confirms distinct channels.
- Teleworkability matters: **ChatGPT coefficient is zero in teleworkable occupations** but significant in non-teleworkable (β̂₂ = −0.233, p < 0.01). AI posting displacement concentrates where remote work cannot cushion the impact.

### Within-employer compositional shift IS AI

Using employer-level difference-in-differences (replaces occupation × time fixed effects with employer × quartile × month FEs), identifying purely from within-employer recomposition across AI-exposure quartiles:

- **22–25 year olds in high-AI-exposure occupations: −5.5% by 2025H1** vs less exposed occupations within the same employer.
- 26–30 year olds: −4.9%.
- Progressively smaller effects for older groups.
- **50+: +1.3%.**
- Pattern monotonic by 2025H1.

### Effect 2x larger for young women

- 22–25 cohort: γ̂₂ = −0.016 (women) vs −0.007 (men).
- Reweighting women's occupational distribution to match men's narrows gap by one-third — remainder reflects within-occupation gender differences.

### Mechanical composition limited

- Only 13% of young entrants first observed in 2023 are in top-quartile AI-exposed occupations — limits scope for mechanical composition channel.
- Pre-trend slopes for 22–25 and 26–30 are *positive* (opposite to post-treatment decline) — works against the finding, strengthening it.
- Placebo test (treatment date = July 2022, between rate hike and ChatGPT) shows flat coefficients through 2023 before sharp acceleration in 2024–2025 — distinct AI signal.

### Cross-country comparison

- Consistent with **Brynjolfsson et al. (2025)** US payroll data finding 16% employment decline among young US workers in AI-exposed occupations.
- Consistent with **Hosseini Maasoum & Lichtinger (2025)** US resume/job-posting data showing seniority gradient.
- **Contrasts with Finnish null** (Kauhanen & Rouvinen 2025, 2026) — same employer-level design, no effect found in Finland.
- Authors note Sweden has accelerating decline; Finland has zero — important national variation.

## Methodology in brief

- **4.6 million job ads** from Platsbanken (Sweden's largest recruitment platform; private/local-government employers were legally required to post there until 2007), Jan 2020 – Feb 2026.
- AI exposure: **Dynamic AI Occupational Exposure (DAIOE) index** (Engberg et al. 2024), maps AI benchmark capabilities to Swedish SSYK 2012 4-digit occupations. Quartile classification (Q4 = most exposed).
- **Population-register employer declaration (AGI)** data from Statistics Sweden — full population of employed individuals with SSYK occupation, age, gender.
- Two-period difference-in-differences using Riksbank rate hike (April 2022) and ChatGPT (Dec 2022) as separate treatment dates.
- Employer × quartile × month fixed effects absorb time-varying employer-level shocks; identification purely from within-employer recomposition across AI-exposure quartiles.

## Implications for AdaptAI

The single most useful **age-gradient evidence** in this batch — translates directly to advice for younger UK workers.

### Calculator (`/calculator`)
- **Strong case for an age-aware insight** on calculator results. Currently the calculator does not ask age and so cannot surface this. Either:
  - Add an optional age question, mapped to an age-band insight (under-30 in high-exposure → strong "canaries" warning).
  - Surface generic "if you are early-career" framing in High-exposure result band.
- **`jobTitlesWeighted` defaults:** for roles where the entry-level weight should differ from the senior weight (Programmer, Software Engineer, Marketing Coordinator, Customer Service Representative), this paper supports differentiating.
- **Methodology modal copy:** the Sweden vs Finland divergence (same design, opposite results) is a useful point about uncertainty — supports the calculator's "guidance not prediction" framing.

### Big Picture (`/big-picture`)
- **`bigPictureData.unemploymentByAge.youth16to24`:** paper provides direct evidence the youth band is the divergence vector. Refresh trajectory and consider adding a "22–25 in AI-exposed occupations: −5.5% by 2025H1" callout.
- **`bigPictureData.unemploymentByAge.mature50plus`:** paper finds **+1.3% gain** for 50+ in high-exposure occupations within the same employers — the opposite direction. Refresh trajectory.
- **`bigPictureData.graduateEmployment`:** the entry-level squeeze is exactly what the graduate-position arrays are tracking. Strong supporting evidence for the `withAI` projection trajectories that show steeper decline than `withoutAI`.
- **`bigPictureData.skillsDemand.growing` and `.declining`:** paper does not contribute directly to skills lists.
- **Narrative copy on `/big-picture`:** add explicit framing "evidence to date suggests AI changes hiring composition (especially for under-30s) more than it changes aggregate demand". This is the central contribution of the paper.

### Industry Insights
- **All 17 industry pages:** the entry-level / senior hire divergence is a cross-cutting story. Industries with high AI exposure (tech, finance, professional services, admin) should foreground the age-gradient finding.
- `/industries/tech` — direct application; software/IT roles are the clearest "canaries" in the US data.
- `/industries/admin-customer-support` — non-teleworkable customer service is exactly where the AI posting decline shows up.

### How to Adapt
- `/adapt/career-transitions` — **load-bearing source for advice to young workers**. The paper's findings are the strongest case for "if you are 22–25 in an AI-exposed role, take the threat seriously and act now". The 50+ gain finding is the right counter-evidence to share with older workers.
- `/adapt/upskilling` — paper does not directly evaluate upskilling effectiveness, but its findings on entry-level squeeze make upskilling more urgent for early-career workers.
- `/adapt/rights-protections` — Sweden vs Finland divergence (same design, opposite outcomes) is evidence that **policy and institutional context matter**. Useful framing for the page.
- **Quick Practical Guides:**
  - `/guides/freelancer` — paper is silent on freelance/self-employment.
  - `/guides/office-worker` — entry-level office workers are the most affected demographic in the paper.
  - `/guides/manager` — supports the framing that managers are seeing teams age up (50+ retained, 22–25 not hired) — relevant to leadership narrative.

## Related notes

- [[Anthropic Observed Exposure]] — US young-worker hiring slowdown signal using a different methodology.
- [[Atlanta Fed CFO Survey 2026]] — independent CFO-reported evidence of within-firm compositional shift.
- [[DSIT UK Labour Market Assessment]] — UK youth employment data (16–24 in programming -44% in 2024) consistent with the canaries finding.

## Caveats & limitations

- **Sweden, not UK.** Direct UK applicability needs verification — Brynjolfsson et al. for US is consistent, but Finland is not.
- DAIOE index based on AI capability benchmarks; capability moves fast — exposure measure may be obsolete in two cycles.
- Pre-trend coefficients are jointly significant — paper itself flags caution. Pre-period magnitudes comparable to average post-ChatGPT coefficient. However, pre-trends for 22–25 and 26–30 are *positive*, working against the finding.
- Employer-level FEs absorb employer-specific shocks — does not capture inter-firm sorting.
- Growth in platform-based self-employment (F-skatt) among young workers may lower measured employment but should not differentially affect AI-exposed occupations.
- 22–25 bracket captures early-career workers (Swedish bachelor's graduates are minimum 22), but international comparison needs care for different graduation timing.
- Effect twice as large for young women — paper does not fully disentangle whether this is driven by occupational sorting, gendered task allocation, or hiring discrimination.