Dissecting Canaries in the Coal Mine

I was reading Canaries in the Coal Mine? from Stanford's Digital Economy Lab recently, and one thing kept bothering me.

The paper is not really telling us the story we have become used to hearing about AI and jobs. It is not a story about mass layoffs or some overnight collapse in employment.

It is a much quieter story.

The bottom of the career ladder may be disappearing.

The Stanford researchers looked at millions of actual payroll records through June 2026 and found that, while employment overall has continued to grow, younger workers in occupations most exposed to AI are increasingly falling behind their peers in less-exposed occupations.

That made me go back to a prediction from almost a decade ago.

In 2016, Geoffrey Hinton made a prediction that sounded outrageous at the time: we should stop training radiologists.

Machines, he argued, were already becoming remarkably good at reading medical scans. Give it five or ten years and it would be obvious that we no longer needed to train so many people to do the job.

There are more radiologists today than there were then. They earn more. And many health systems still say they do not have enough of them.

At first glance, this seems like another reminder that predictions about AI and employment tend to be wrong.

But there is a more interesting lesson here.

Reading a scan is only one part of a radiologist's job. A radiologist interprets findings, talks to clinicians, makes decisions, handles uncertainty, documents cases and takes responsibility for what happens next.

Make one of those tasks dramatically cheaper and you do not necessarily eliminate the profession. Sometimes you make the entire service cheaper and more accessible, which increases demand for everything else.

That is the optimistic side of the AI story, and there is plenty of evidence for it.

The problem is that this logic does not apply everywhere.

And that is what makes this Stanford paper so interesting.

Read the paper here: Canaries in the Coal Mine?, Stanford Digital Economy Lab (PDF)

What the data actually says

The August 2026 update to Canaries in the Coal Mine?, by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford's Digital Economy Lab, looks at payroll records from ADP covering millions of US workers through June 2026.

This matters because we are not talking about surveys asking people whether they feel AI is changing work.

We are looking at actual employment records.

And the first finding is important precisely because it is not particularly dramatic.

There has been no mass employment collapse.

Total employment in the sample grew by roughly 6% since November 2022. Employment in the fifth of occupations most exposed to AI grew by about 4%.

That is a slowdown. It is not an apocalypse.

But then you look at younger workers.

Among people aged 22 to 25 working in the most AI-exposed occupations, employment is roughly 19% below where it would have been if it had continued to track employment among similarly aged workers in less-exposed occupations.

That gap was around 15% a year earlier.

It has continued to widen.

Put another way, employment for 22-to-25-year-olds in the two most AI-exposed groups fell about 11% between November 2022 and June 2026.

For the same age group in the three least-exposed groups, employment increased by about 10%.

And there is another detail that may be even more important.

The same pattern does not appear among workers over 30 in those occupations.

The issue, at least so far, is not that AI is simply eliminating jobs across the board.

It is affecting who gets hired into them.

Nobody is getting fired. The job just never appears.

This is the part that I think deserves far more attention.

The data suggests the employment effect for young workers is showing up through hiring rather than layoffs.

Separation rates for younger workers in exposed occupations actually fell, rather than rising dramatically.

People are not necessarily being marched out of offices or call centres.

Instead, some of the roles that would previously have been posted are simply not being opened.

That distinction is easy to miss.

Layoffs make headlines. Companies announce them. Employees talk about them. The media reports them.

A job that never gets advertised does not generate a press release.

That makes this kind of change almost invisible until you start looking at the numbers.

And it changes the question we should be asking.

The question is not simply:

How many jobs is AI destroying?

It is:

Which jobs are companies choosing not to create anymore?

The most important distinction: automate or complement?

The Stanford research becomes especially interesting when it separates AI exposure into two different forms.

The difference is enormous.

In occupations where AI use leans toward automation, employment for the youngest workers falls substantially. The effect gets progressively smaller as workers get older and is no longer statistically distinguishable from zero in older age groups.

In occupations where AI is used more as a complement to people, that pattern disappears.

In fact, the occupations with the strongest complementarity have some of the fastest employment growth among younger workers. The relationship is also positive for experienced workers aged 41 to 49.

That is a much more useful way to think about AI and employment than the usual argument about whether AI is good or bad for jobs.

Because it suggests something uncomfortable:

The employment impact of AI is not entirely something that happens to organisations.

Some of it is something leaders decide.

Those are different choices, and they produce different labour-market outcomes.

Brynjolfsson often talks about AI as a way of amplifying intention.

I think that framing is useful.

Contact centres are where this gets real

If you want to see the transition happening in front of you, contact centres are probably one of the clearest examples.

Brynjolfsson's earlier research on a large call centre showed AI primarily helping human agents. The technology helped agents answer questions faster while the human still owned the conversation.

That is augmentation.

Today, some of the same organisations are moving further.

A meaningful share of customer interactions can now be handled directly by AI agents.

That changes the economics.

You do not need to imagine a future where every worker is replaced overnight.

The company can simply decide that it needs fewer people to handle the same volume of demand.

That is a much quieter transition.

Goldman Sachs has estimated that US call-centre employment is around 39% below trend, making it one of the most disrupted occupational categories in its analysis. PwC's 2026 Global AI Jobs Barometer similarly finds that junior roles in AI-exposed areas are increasingly being asked to demonstrate skills such as judgement, leadership and strategy that were traditionally developed later in a career.

And that creates a strange paradox.

The entry-level job does not necessarily disappear.

Instead, the entry-level job becomes harder to qualify for.

The first rung gets higher.

The Stanford paper is careful about what it does not prove

One thing I appreciate about the Stanford work is that it does not pretend the data proves more than it actually does.

There are legitimate alternative explanations.

AI-exposed occupations were already diverging from other occupations before ChatGPT arrived, especially during and after COVID.

Education is another confounding factor. Highly AI-exposed occupations are also disproportionately occupations requiring degrees, and once education is controlled for, the estimated employment effect becomes significantly smaller.

There is also a serious argument that interest rates played a bigger role than AI.

The Economic Innovation Group has made that case, pointing to the timing of the labour-market slowdown and the rapid rise in rates in 2022.

That is a reasonable challenge.

And this is where the evidence becomes interesting again.

The employment gap did not disappear when interest rates stabilised and later declined. Instead, it continued widening through mid-2026.

The authors also find that controlling for occupational exposure to interest rates changes the estimate by less than a percentage point.

Remote work has been proposed as another explanation, but that does not fully account for the pattern either.

None of this turns the research into causal proof.

It does something more modest.

It shows us a pattern that is difficult to ignore.

And sometimes patterns are where the important questions begin.

The real problem is the pyramid

This is the part that concerns me most.

Most professional organisations are pyramids.

There is usually a broad base of people doing relatively routine, checkable work.

Eventually, some become the people managing teams, clients and complicated problems.

That is how organisations manufacture experienced talent.

Except AI is increasingly very good at the work that used to make up the bottom of the pyramid.

And that creates a problem.

Imagine every company independently decides to hire fewer graduates because AI can already do much of the junior work.

For an individual company, that decision can make perfect financial sense.

Nothing immediately breaks.

But if everyone does it, something happens five years later.

Brynjolfsson describes this as a coordination problem. I think that is exactly right.

Every individual company has an incentive to reduce the cost of developing junior talent.

But the industry as a whole still needs people to become seniors.

Someone has to create them.

One interesting answer comes from Infosys

There is at least one approach that deserves more attention.

Brynjolfsson points to Infosys as an example of a company that has continued hiring graduates while changing what those graduates are trained to do.

Instead of putting young employees through the same routine work that AI agents can now handle, the company is placing more emphasis on skills such as framing problems and exercising judgement.

In other words, they are making the apprenticeship explicit.

That idea matters because the old apprenticeship model depended heavily on proximity.

Gradually, almost without noticing, become good.

AI removes a lot of the routine work that made this process possible.

So organisations may need to redesign how people learn instead of assuming it will happen naturally.

The problem is that this investment costs money today.

The payoff might not arrive until 2031.

That makes it very easy to postpone.

What this means for workforce planning

I think four practical lessons follow from the evidence.

1. Ask whether every AI deployment is automating or complementing

This should be a basic design question before anyone talks about ROI.

The distinction matters because the employment outcomes look very different.

2. Stop thinking of entry-level headcount as just a cost

Junior employees are not only today's productivity.

They are tomorrow's experienced workforce.

Eliminating those roles may improve a budget this year while creating a talent shortage several years from now.

That is a delayed cost, which makes it dangerously easy to ignore.

3. Make the apprenticeship deliberate

If AI takes away the routine tasks through which junior employees historically learned, companies have to replace that learning mechanism.

Do not assume that sitting near a senior person will magically produce judgement.

4. Hire for the things AI struggles to learn from documentation alone

The Stanford research points towards an important distinction between codified knowledge and tacit knowledge.

AI is naturally strong at structured, documented, teachable work.

Human experience becomes more valuable when the job depends on judgement, context, ambiguity and accumulated practice.

That should influence how organisations think about hiring and development.

The next phase of work may look very different

There is a broader shift hiding underneath all of this.

Brynjolfsson has a useful way of breaking work into three parts:

AI is getting increasingly good at the middle.

That means value starts moving towards the first and third.

The person who knows exactly what needs to be done becomes more important.

So does the person who can tell whether the AI did it properly.

That changes what management itself looks like.

A growing number of people may spend less time doing individual pieces of work and more time managing a fleet of AI systems: defining the objective, reviewing outputs, correcting mistakes and deciding what happens next.

We have not traditionally taught twenty-two-year-olds how to do that.

We are going to have to.

The honest ending

There are two very different futures sitting behind this research.

The optimistic one is extraordinary.

AI could create enormous productivity gains, increase living standards and accelerate progress in areas such as medicine and science.

The pessimistic one is equally serious.

AI could concentrate wealth and power, distort information ecosystems, enable dangerous technologies and deepen inequality.

Neither outcome is inevitable.

And that may be the most important point.

More powerful tools do not automatically mean less human agency.

They can mean more.

The real question is how we choose to use them.

That choice is not being made only in government offices or research labs.

It is being made inside workforce plans, operating models, budgets and technology projects every day.

The question is what happens when an entire industry makes the first choice at the same time.

Because maybe the biggest AI employment story is not that the ladder is being kicked away.

Maybe it is that we are quietly removing the bottom rung.

And by the time we realise what it was holding up, rebuilding it may be much harder than we expected.