Three years in: no disruption, and one 20% hole

TL;DR. The Budget Lab at Yale examined the first 33 months after ChatGPT and found no detectable economy-wide disruption: occupational and industry mix flat or within historical ranges, and the pace of change comparable to personal computers in 1984 and the internet in 1996. Danish administrative records across eleven exposed occupations found essentially zero effect on earnings or hours through 2024. A US survey finding 35.9% generative AI use by December 2025 reported small positive wage effects and no significant employment declines. And the Stanford AI Index, using payroll data across millions of workers, reports employment for software developers aged 22 to 25 down nearly 20% since late 2022 while older developers at the same firms grew 6 to 12%. Both findings are well evidenced. The disagreement is entirely about scope.

---

Status: contested, and the contest is legitimate. Primary sources: the Budget Lab at Yale's CPS analyses; Humlum and Vestergaard on Danish administrative data; Hartley and colleagues on US survey data; and the Stanford HAI AI Index payroll analysis. Disclosure: one of the companies whose usage data the Yale team incorporates is Anthropic, which makes the model used in drafting parts of this site. No finding here is presented as favouring any party.

---

What the aggregate measures say

The Budget Lab at Yale published in October 2025 on the first 33 months since ChatGPT's launch. Using Current Population Survey data, it tracked occupational and industry mix against historical benchmarks with a dissimilarity index.

The mix barely moved. Occupational dissimilarity, industry dissimilarity and exposure metrics all sat flat or within historical ranges. Employment shifts followed long-running trends: clerical decline, service-sector growth.

And the comparison the study makes is the useful part. It placed current change against the personal computer wave from 1984 and the internet wave from 1996. Today's changes are unfolding at a similar pace, which is neither reassuring nor alarming and is a fact worth holding.

Its subsequent update was blunter still. Measures of exposure, automation and augmentation show no sign of being related to changes in employment or unemployment, and better data is needed.

Two independent studies agree. Humlum and Vestergaard linked survey-reported ChatGPT use to Danish administrative records across eleven exposed occupations and found essentially zero effects on earnings or hours through 2024. Hartley and colleagues found 35.9% of US workers using generative AI by December 2025, with small positive wage effects and no statistically significant declines in job openings or employment in exposed occupations.

Three datasets, three methods, three countries' worth of data, one answer: continuity.

And what the age gradient says

The Stanford AI Index drew on payroll records covering millions of workers across tens of thousands of firms from 2021 to 2025.

Employment for software developers aged 22 to 25 fell nearly 20% since late 2022. Employment for older developers at the same firms grew 6 to 12%.

The phrase "at the same firms" is what makes this hard to dismiss. A within-firm comparison controls for the thing that usually explains employment changes: a firm hiring fewer juniors because demand fell would also be hiring fewer seniors. These firms did the opposite.

And software is the most AI-exposed occupation by essentially every exposure measure, which is where a first effect would be expected to appear.

Related work reaches similar conclusions. Brynjolfsson, Chandar and Chen documented six facts about early employment effects concentrated in entry-level segments of highly exposed occupations, and other researchers report pressure among younger workers and new hires.

Why both can be right

They measure different things, and the difference is the entire story.

The aggregate studies ask whether the economy's occupational structure has shifted. It has not. Software developers are a small share of employment, and a 20% fall in one age band of one occupation is invisible in a national occupational mix.

The payroll study asks whether a specific cohort in a specific occupation is being hired. It is not, relative to its older colleagues at the same employers.

Neither answer contradicts the other. They are answers to different questions, which is the scope problem in its most consequential form so far in this corpus: a genuine effect on the people it lands on, invisible at the level most policy discussion operates.

What the careful researchers say about the careful finding

The Budget Lab looked at the same age question and found mixed evidence.

Its own signal, that occupational-mix dissimilarity between workers aged 20 to 24 and 25 to 34 has risen slightly faster and sits at the high end of the historical range, is described by its authors as a nascent signal, with small samples, and with the observed trend possibly predating ChatGPT.

They explicitly decline to characterise it as confirmed AI-driven displacement.

The Budget Lab also flags a different pattern in early 2026: low layoffs alongside low hiring, particularly low hiring of unemployed workers. It does not attribute this to AI, and it is exactly the macroeconomic condition under which entry-level cohorts suffer most regardless of technology.

That is the confound, and it is not resolved. Interest rates, a post-2021 correction in technology hiring, and a low-flow labour market all predict fewer junior developers. AI predicts the same thing. Nobody has separated them cleanly, and the within-firm design, while strong, does not do it either: a firm can freeze junior hiring for budget reasons while retaining seniors it cannot replace.

Three things this establishes

Aggregate stability and cohort damage are compatible. A finding of no economy-wide disruption is not a finding that nobody was affected, and it is routinely reported as though it were.

The pace comparison deserves more attention than it gets. Change at the rate of the personal computer and internet waves is substantial change. The PC wave transformed clerical work over two decades. Reading "similar to previous waves" as reassurance requires forgetting what the previous waves did.

And the strongest available design still cannot separate AI from the cycle. A within-firm age comparison controls for firm-level demand and does not control for a firm's decision to stop hiring juniors for reasons unrelated to capability. The honest position is that something is happening to the bottom rung and its cause is not established.

What it does not establish

That AI has not displaced anyone. Aggregate null results bound the size of the effect at the economy level; they say nothing about individuals.

That the entry-level effect is AI. The Budget Lab's caution is well founded and the confounds are severe.

That the aggregate will stay flat. Three years is early. The studies say so themselves, and the previous waves the comparison invokes took decades to work through.

And nothing about which jobs are next. Exposure measures predict where effects would appear if they appear, and have so far predicted the location of a signal rather than its magnitude.

What is unresolved

Whether the entry-level effect persists or reverses. If it is a hiring-cycle artefact it should reverse; if it is structural it should spread to adjacent occupations.

What happens to the skill pipeline. If junior roles are how seniors are produced, a sustained gap has effects that arrive years later and are invisible in current employment data.

Whether better data resolves it. The Budget Lab's repeated conclusion is that better data is needed, which is unusual candour and also a statement that current instruments cannot answer the question.

And whether the low-flow labour market is itself an AI effect. Low layoffs with low hiring could reflect employers uncertain about future staffing needs, which would make the macro condition partly downstream of the technology rather than a confound to be removed.

The counter-argument

Treating aggregate nulls and a cohort finding as equally weighted is generous to the cohort finding. One is replicated across three methods and two countries. The other is a single payroll dataset in a single occupation, and the researchers closest to the aggregate data looked at the same question and found mixed evidence.

The within-firm design is weaker than it appears. Firms cut junior hiring first in every downturn, for reasons of training cost and immediate productivity, and 2022 to 2025 was a severe correction in technology employment. The pattern is exactly what a hiring freeze produces.

The pace comparison cuts both ways. If AI is following the PC and internet trajectories, then the current absence of disruption tells us almost nothing, because those effects were also invisible three years in. That is an argument for concern, not for calm, and this article's framing of it as neutral is a choice.

And the aggregate studies may be measuring the wrong thing. Occupational mix is a coarse instrument. Task composition within occupations can change completely while the occupational label persists, which is what augmentation looks like, and no dissimilarity index would detect it.

The short version

The Budget Lab at Yale examined 33 months after ChatGPT and found no detectable economy-wide disruption: occupational and industry mix flat or within historical ranges, and change unfolding at a pace comparable to personal computers in 1984 and the internet in 1996. Its update found exposure, automation and augmentation measures unrelated to employment or unemployment. Danish administrative records across eleven exposed occupations found essentially zero effect on earnings or hours. A US study with 35.9% AI usage found small positive wage effects and no significant employment declines.

And the Stanford AI Index, on payroll data across millions of workers and tens of thousands of firms, reports software developers aged 22 to 25 down nearly 20% since late 2022 while older developers at the same firms grew 6 to 12%.

Both are well evidenced, and they answer different questions. A 20% fall in one age band of one occupation is invisible in a national occupational mix. Aggregate stability and cohort damage are compatible, and the first is routinely reported as though it settled the second.

The confound is unresolved. The Budget Lab, looking at the same age question, found mixed evidence, called its own signal nascent, noted small samples, and observed the trend may predate ChatGPT. It separately flags low layoffs with low hiring in early 2026, which it does not attribute to AI and which predicts exactly this pattern on its own.

Even the within-firm design does not settle it, because a firm can freeze junior hiring for budget reasons while retaining seniors it cannot replace. Something is happening to the bottom rung. Its cause is not established.

Common questions

Has AI displaced workers at scale? Not on any aggregate measure so far. The Budget Lab at Yale found no detectable economy-wide disruption across the first 33 months after ChatGPT, with occupational and industry mix flat or within historical ranges. Danish administrative records across eleven exposed occupations found essentially zero effect on earnings or hours through 2024, and a US study finding 35.9% generative AI usage by December 2025 reported small positive wage effects with no significant employment declines.

What is the contrary evidence? The Stanford AI Index, drawing on payroll data across millions of workers and tens of thousands of firms from 2021 to 2025, reports employment for software developers aged 22 to 25 down nearly 20% since late 2022, while employment for older developers at the same firms grew 6 to 12%. Related work by Brynjolfsson, Chandar and Chen documents pressure concentrated in entry-level segments of highly exposed occupations.

How can both be true? Because they measure different things. The aggregate studies ask whether the economy's occupational structure has shifted, and it has not. The payroll study asks whether a specific cohort in a specific occupation is being hired, and relative to older colleagues at the same employers it is not. Software developers are a small share of total employment, so a 20% fall in one age band is invisible in a national occupational mix. Aggregate stability and cohort damage are compatible.

Is the entry-level effect definitely AI? No, and the researchers closest to the aggregate data are careful about this. The Budget Lab investigated the age question and found mixed evidence, described its own signal as nascent, noted small sample sizes, and observed the trend may predate ChatGPT. It separately flags a low-layoff, low-hiring labour market in early 2026 which it does not attribute to AI and which produces the same pattern. Interest rates and a post-2021 technology hiring correction are further candidate causes.

Does the within-firm comparison settle it? It is the strongest design available and it does not settle it. Comparing age bands within the same firms controls for firm-level demand shocks, which is why the finding is hard to dismiss. It does not control for a firm freezing junior hiring on budget or training-cost grounds while retaining seniors it cannot easily replace, which is standard behaviour in a downturn and describes 2022 to 2025 in technology employment.

What does the comparison to previous technology waves mean? The Budget Lab placed current change against the personal computer wave from 1984 and the internet wave from 1996, and found today's changes unfolding at a similar pace. That is neither reassurance nor alarm. Those waves transformed clerical and information work over two decades, and their effects were also largely invisible three years in, which means the current absence of aggregate disruption carries less information than it appears to.

What is the most important unresolved question? Whether the entry-level gap persists. If it is a hiring-cycle artefact it should reverse as conditions ease; if it is structural it should spread to adjacent occupations. There is also a delayed consequence nobody can currently measure: if junior roles are how senior workers are produced, a sustained gap has effects that arrive years later and are absent from every present-day employment series.

Why does the framing of these findings matter so much? Because an aggregate null is routinely reported as evidence that nobody is being affected, and it is not that. It bounds the size of the effect at the economy level and says nothing about individuals or cohorts. A study finding no economy-wide disruption and a study finding a fifth of an entry-level cohort gone are both accurate, and only one of them describes what happened to the people in it.