The Stanford data everyone's citing on AI and jobs, and what it actually shows

[LuckySentinel]

Every recent debate about whether AI helps or replaces workers eventually traces back to the same source, the Stanford Digital Economy Lab's Canaries dashboard, built with ADP payroll data covering roughly one in six American workers, tracking employment outcomes by age across occupations most and least exposed to AI

At the headline level, the picture looks calm, across all workers, the most AI-exposed occupations actually grew employment slightly faster since ChatGPT's late 2022 launch than the least exposed ones did. Cut the data by age and career stage though, and a different story appears. For workers aged 22 to 25 specifically, employment in the most AI-exposed occupations, software development, customer service, junior analysts, accounting, is now shrinking at roughly 3.8 percent per year, and that decline has accelerated over time, from about 2.8 percent through April 2024 to more than 4 percent annually since. Every age group 31 and older in those same high-exposure occupations actually grew over the same period, while the 26 to 30 age bracket sits roughly flat

The Stanford researchers, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, offer a clean explanation for why the youngest workers specifically bear the brunt, entry-level work leans heavily on codified knowledge, the kind of information you can write down as explicit rules and learn from a manual, which is exactly the kind of task current AI models handle well. Experienced workers instead lean on tacit knowledge, judgment and context built up on the job over years, something AI still handles poorly. The junior tasks are the ones getting automated, which means the junior rungs of the career ladder are the ones being removed first

Not every economist accepts the AI causation story though. Google's own economists have pointed to interest rate effects instead, and other researchers cite tech sector overhiring during the pandemic, remote work distortions and general post pandemic noise as competing explanations. The Yale Budget Lab has found no clear employment link through parts of 2025, and a New York Fed study of job postings found the relative slide in AI-exposed roles actually started before ChatGPT even shipped. The researchers themselves call the pattern canaries in the coal mine deliberately, an early warning signal rather than a final verdict, meaning the youngest workers are where any real effect would show up first, and right now, on this specific dataset, they're the ones showing it

Will your job be replaced by AI? Here are the roles most affectedfinance.yahoo.com

[Orbit William]

The distinction between codified knowledge that AI handles well and tacit knowledge that only comes from years on the job is the clearest explanation I've read for exactly why the youngest workers specifically are the ones losing ground

[SilverSurfer51]

The competing explanations, interest rates, pandemic overhiring, remote work distortions, deserve to be taken seriously rather than dismissed, correlation with ChatGPT's launch timing isn't automatically proof of causation

[ShawnMichaels_99]

Every age group 31 and older actually growing in the same high exposure occupations while 22 to 25 year olds shrink is such a stark divergence within the exact same job categories, that's not a story about occupations disappearing, it's about who gets to keep doing them

[Harry64]

Canaries in the coal mine as a deliberate framing choice by the researchers themselves is a smart, honest way to present genuinely uncertain but concerning data without overclaiming

[EventHorizon55]

The acceleration from 2.8 to over 4 percent annual decline is the part that would worry me most if I were currently job hunting at 23, that's not a one time adjustment, that's a trend still getting worse

[Molly76]

This ties directly into the apprenticeship pipeline concern that's come up in insurance, law, and now general early career hiring, strip out the entry level tasks and you risk losing the mechanism that trains the next generation of experienced workers entirely

[Seb83]

The frustrating part is how often people quote a single statistic without the surrounding context.

A decline in employment for one group does not automatically mean AI caused every bit of it.

Economic cycles, hiring freezes, interest rates, and company restructuring can all overlap.

That is why longitudinal datasets are useful. They let you look for patterns instead of grabbing one dramatic headline.

Still, if the decline is accelerating, that deserves attention rather than dismissal.

Nobody benefits from pretending everything is fine when the data suggests otherwise.

[Matthew97]

One thing that stood out to me is that entry-level roles seem to be getting squeezed from both directions.

Companies want experienced people because they need less supervision.

At the same time, AI makes experienced employees more productive, reducing the pressure to hire juniors.

That combination could make getting the first job much harder than keeping the tenth.

It is a bit of a catch-22 for graduates.

[AlphaOscar89]

There is a danger in treating AI as either a miracle or a disaster.

Reality is usually somewhere in the middle.

Spreadsheets eliminated plenty of bookkeeping work, but they also created entirely new jobs around finance and analysis.

AI may follow a similar path, although the transition period could be much rougher.

The people caught in that transition are the ones policymakers should be thinking about.

[NightCrawler33]

Something else worth mentioning is that percentages can sound scarier than absolute numbers.

A four percent annual decline is significant, but it still needs to be translated into actual jobs and industries before drawing sweeping conclusions.

Different sectors are moving at very different speeds.

Software, customer support, marketing, and legal research are not experiencing the same pressures.

Blanket statements rarely survive a closer look.

[SortedIt]

One possible response is changing how education works.

Universities and colleges may need to spend less time teaching routine production tasks and more time teaching evaluation, communication, critical thinking, and domain expertise.

Those are the areas where people still add the most value.

Knowing how to work with AI could become just as important as knowing how to use spreadsheets became twenty years ago.

That seems like a more productive response than trying to ignore the technology.

[NightOwl83]

There is a weird irony here.

Young workers are usually the quickest to adopt new technology, yet they may also be the group most affected if entry-level work becomes easier to automate.

Hopefully businesses do not lose sight of the value of developing talent.

Saving money this quarter is one thing.

Ending up with a shortage of experienced professionals in a decade would be a much bigger problem.

[Elk31]

A little skepticism is healthy whenever everyone starts citing the same research.

Replication matters.

If similar conclusions start appearing from different countries using different datasets, confidence in the trend grows.

One influential report is valuable, but it should not become the only reference anyone ever mentions.

Science works best when multiple sources point in the same direction.

[Tracey]

The first jobs people get are often where they make mistakes and learn from them.

If AI removes too many of those junior opportunities, where does the next generation of experienced workers come from?

That feels like the long-term issue hiding underneath the employment numbers.

Businesses still need senior staff ten years from now.

Those people have to start somewhere.

[Scholar29]

The conversation also needs to include wage growth instead of only employment.

Someone might keep their job but find salary increases slowing because AI expands the available labor pool.

That is a different kind of pressure that does not show up in simple employment charts.

People tend to focus on layoffs because they are dramatic.

Pay trends can quietly reshape an industry over several years.

[EventHorizon63]

Companies are probably making hiring decisions differently already.

Instead of asking whether one person can write reports all day, they may ask whether one person plus AI can handle the workload of three people.

That changes workforce planning even if no jobs disappear overnight.

Productivity gains often arrive gradually rather than with one giant announcement.

That is why these datasets are so interesting.

[BinaryMonk78]

The acceleration is what caught my attention too.

Small year-to-year changes can be dismissed as noise.

When the rate itself starts increasing, it becomes much harder to ignore.

That does not automatically prove AI is the only cause, but it certainly makes the hypothesis worth investigating much more seriously.

Good data should drive better questions rather than immediate conclusions.