September 2026 · Working notes
Where is AI showing up in the data?
Since ChatGPT arrived in late 2022, the question I hear most is whether generative AI is already changing the economy. I have looked for it in a dozen places: mortgage and business credit, jobs and wages, job vacancies in three countries, how people spend their time, mental-health care, marriage and births, admission to university programmes, and industry prices. The short version: one signal keeps showing up, in mortgage refinancing. Almost everywhere else, the honest answer is not yet.
In this note: Measuring exposure · The mortgage signal · Where else I looked · Why “not yet” matters · What’s next
Measuring exposure
Everything below starts from an AI-exposure score for each occupation: roughly, how much the abilities a job requires overlap with what current AI, and language models in particular, can do. I aggregate these scores to industries and to US counties in two ways — by where people work (the county's 2019 industry mix) and by where people live (the 2019 occupations of employed residents). The two are less alike than one might expect: across counties they correlate at only about 0.26.
Exposure is not adoption, and it is not job loss; it measures where AI could matter. As a check, I compare it with what people actually do with AI, using Google's occupation-level Gemini usage data and Anthropic's Claude usage data. Those usage data were collected in 2026, after most of the outcomes below, so they help validate the measure rather than prove any effect.
The one signal that keeps showing up: mortgage refinancing
The US mortgage data (HMDA) record home-loan applications lender by lender and county by county. Comparing 2019 with 2023–2025 across about 1.6 million lender-county-year cells, more AI-exposed places see less mortgage lending. That broad pattern is not, by itself, an AI effect: high-exposure counties are large, tech-heavy, remote-work-intensive places whose mortgage markets move with the rate cycle. Once I control for rate sensitivity, local geography and proxies for household balance sheets, the broad gap largely disappears.
What survives is a sharper comparison. Within the same county and year, and for the same lender, refinance lending falls more than purchase lending in more AI-exposed counties: about 14% fewer refinance loans, relative to purchase loans, per standard deviation of exposure (a Poisson estimate; the 95% interval is roughly 10–18%). Three things make me take this seriously:
- It survives proxies for the obvious confounder. Controlling for interest-rate lock-in (each county's 2020–21 refinancing patterns, plus machine-learning predictions of its rate sensitivity) shrinks the refinance gap by about a third on the transformed-count scale, but it stays clearly negative.
- It holds up on revised data. Re-estimating the core specification on the revised 2024 HMDA file and the 2025 snapshot leaves the estimate essentially unchanged.
- Where borrowers live matters. In separate regressions where both measures enter, the fall in mortgage lending loads mainly on residents' exposure — the people whose income repays the loan — rather than on local employers'. A resident measure built from Google's 2026 usage data also predicts lower lending, though only as a check after the fact.
Show the numbers
| Year | Estimate | 95% interval |
|---|---|---|
| 2019 | 0 (reference) | – |
| 2020 | +0.082 | +0.066 to +0.098 |
| 2021 | +0.028 | +0.010 to +0.046 |
| 2022 | −0.082 | −0.102 to −0.062 |
| 2023 | −0.153 | −0.181 to −0.126 |
| 2024 | −0.125 | −0.148 to −0.101 |
| 2025 | −0.092 | −0.114 to −0.069 |
There are also clear limits to what these data can say. Purchase lending barely moves, so this is not a credit crunch. HMDA cannot tell lenders screening more carefully apart from borrowers choosing not to refinance. A narrower story I initially expected — lenders pulling back from large loans just above the conforming limit (the ceiling for loans Fannie Mae and Freddie Mac can buy) — did not hold up. And early changes in credit do not reliably predict later unemployment once differences across states are accounted for. The finding is a persistent reallocation of refinancing, not yet evidence that lenders are pricing AI risk.
Where else I looked
| Where I looked | What I found | Reading |
|---|---|---|
| Mortgage refinancingHMDA, 2019 vs 2023–25 | Refinancing falls about 14% relative to purchases per SD of exposure, within lender and county. | Persistent |
| Small-business loansCRA bank lending | No fall in total lending; the mix shifts toward firms with revenue of $1 million or less. | Mix, not volume |
| Corporate loansSyndicated loans, 10-K text | Tighter terms at the industry level fade when comparing the same borrower. Firms whose earlier 10-Ks framed AI as a risk pay slightly higher commitment fees, in a small sample. | Conditional |
| Jobs and businessesCounty employment, 2024 Q1 to 2026 Q1 | No detectable change in employment or establishments; large declines are ruled out. | Not yet |
| WagesIndustry, occupation and county data | Industry weekly wages about 1.4% lower per SD after allowing for trends; shorter-window checks are weaker, and the county follow-up has the opposite sign. | Mixed |
| UnemploymentCounties, mid-2025 to mid-2026 | Where residents are more exposed, unemployment rises slightly: about 0.02 percentage points per SD, comparing counties within the same state. | Small |
| Job vacanciesUS, Canada, UK | The countries disagree: more openings per hire in the US, more vacancies but lower offered wages in Canada, and somewhat fewer vacancies in the UK in some specifications. | Mixed |
| Time useAmerican Time Use Survey, to 2025 | No detectable change in work, sleep or care time; about 2.7 fewer minutes a day of leisure computer use per SD. | Exploratory |
| Mental-health careNational Health Interview Survey | Therapy use in the past year up about 1.1 percentage points per SD (2024 vs 2021) — which may reflect seeking care as much as distress. | Exploratory |
| Marriage and birthsAmerican Community Survey | High-exposure places have lower birth rates, but that gap was already there; changes since 2022 show no pattern, and neither do marriage or divorce. | Not yet |
| University admissionsFrance and Turkey | An early sign that exposed programmes in Turkey became less selective in one track (and more selective in another) disappears with a finer occupation mapping; no broad avoidance in France. | Fragile |
| Industry pricesProducer prices, 105 industries | No clear or robust relation between exposure and 2022–25 producer-price inflation, though moderate effects can’t be ruled out. | Not yet |
Why “not yet” is still informative
AI can change expected risk much faster than it changes jobs. Tasks can be reorganised in months, while employment, firm counts and household income move slowly — and credit markets can, in principle, respond to expectations before any of those move. The evidence so far fits that timing, though it cannot show what lenders expected: a signal in refinancing, and little yet in employment, firms or prices. Some of the nulls are also tight enough to be informative in their own right. For county employment growth between early 2024 and early 2026, the estimates rule out declines larger than about half a percent per standard deviation of exposure.
A second lesson concerns the exposure scores themselves. A single score treats every exposed task alike, but in a task-based model where AI replaces some work and creates other work inside firms, what matters for wages is which tasks are exposed. Calibrating that model to US industries with O*NET task content and OEWS employment, a single score and the full task profile disagree on the sign of the predicted wage effect in about one industry in nine today, and in roughly 30% of cases in stylised scenarios where AI capabilities become more task-specific. For now, the extra task detail improves the fit to observed industry wages only modestly, and not significantly. If capabilities do diverge, though, a single exposure number would become a less reliable guide to who is affected.
What I am watching next
- Loan performance. Whether the refinancing shift anticipates later delinquencies — the test that would separate careful screening from realised risk.
- University placements. Turkey's 2026 placement results, to see whether early shifts in the selectivity of exposed programmes settle into a pattern.
- Usage data. New releases from AI developers that measure adoption directly, rather than exposure.
These are working notes, not a paper. Estimates are per standard deviation of AI exposure and are associations, not causal effects. Detailed tables are available on request.