What the Hiring Numbers Can Tell Us

A person who has sent applications for weeks does not need a chart to tell them that finding work can be hard. They need a chart to answer a different question: is their experience part of a wider change, and what sort of change is it?

That second question deserves care. If we misread the evidence, we can give the person poor advice and ask institutions to solve the wrong problem. We might tell someone to buy more training when employers have reduced hiring. We might blame every rejection on technology when a particular qualification, location or work sample could make a difference. We might mistake a broad improvement for proof that no group is being squeezed.

The evidence on AI and early work is important, unsettled and varied. Its differences are part of what we need to understand.

Begin with five separate things: the tasks a technology could affect, the jobs employers advertise, the people they hire, the people already employed and the abilities people acquire. A study of one can raise questions about the others. It does not measure them all.

The door and the room

Employment is the number of people in work at a given time. Hiring is a flow of people into jobs. Separations are a flow out. A room can remain fairly full while its door becomes harder to enter.

Take a deliberately simple example. A team begins with twenty workers. During a period, two leave and two are hired. It ends with twenty. In the next period, no one leaves and no one is hired. It still has twenty workers, but there was no new chance to join.

This example is arithmetic, not a description of any particular company. It shows why an applicant and an existing employee can experience the same workplace differently. The applicant feels the absence of openings before the employment total reveals much change.

Now add a tool that helps the existing team do more. The employer might expand and hire. It might keep headcount steady while taking on more business. It might replace a departing worker’s tasks with software. Or it might reduce staff. Which outcome occurs depends on demand, costs, what the tool can do, and decisions inside the firm.

A demonstration of faster work cannot tell us which employment outcome will follow. To learn that, we need evidence about employment decisions.

The same caution applies to job advertisements. An advertisement may represent an intended opening, a replacement, a repeated posting or a search that is later abandoned. It is useful evidence of advertised demand. A completed hire is another event. We will use a posting analysis later to examine what kinds of opportunities are being offered, while keeping that limit in view.

A serious warning from young-worker records

A Stanford study by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen follows an ADP employer panel through June 2026. Among workers aged 22–25, employment in more AI-exposed occupations fell behind less-exposed occupations. The authors describe a 19 percent shortfall relative to keeping pace, linked mainly to slower hiring. This is a descriptive comparison, not a count of jobs proved to have been destroyed by AI. Education adjustments reduce the estimates; some trends began earlier, and the employer sample differs from national surveys. Age also does not tell us whether someone is a career beginner.¹

That finding is worth attention without being turned into a verdict about every newcomer. A warning does not have to answer every causal question before we investigate it.

For an employer, the practical question is whether entry hiring has changed and why. Did the team stop hiring for tasks that now receive more software help? Was demand already falling? Did a budget cut affect all recruitment? Were senior roles protected while training places disappeared? Those questions can be answered more usefully inside a firm than by repeating a headline.

For a college, the question is whether its students still have routes into supervised paid work. If employers are changing what they buy from graduates, the college needs to understand the change. A new course title is not enough. Someone has to examine the actual tasks, the entry requirements and the relationship with employers.

For a newcomer, the warning is a reason to look closely at the route you are considering. It is not evidence that your effort has become pointless. Different kinds of work, places and organizations can move in different directions. The value of research is to improve the questions you ask, rather than make one national story stand in for your whole future.

Evidence that complicates the easy explanation

Anders Humlum and Emilie Vestergaard linked Danish AI-adoption surveys with administrative records through December 2024. Their comparisons found no detectable differential effects on average earnings or recorded hours, and no relative fall in early-career employment shares at adopting workplaces. They did observe wider changes in exposed occupations, but their adoption comparison did not explain those declines. This was not randomized adoption, and the setting, period and measurements differ from the U.S. study.²

It would be convenient if one paper made the other disappear. Research rarely works that way.

Suppose one study asks whether workers in occupations exposed to a technology have followed different employment paths. Another asks whether workplaces reporting adoption have changed differently from those reporting less adoption. These questions overlap, but they are not identical. Firms may anticipate a technology before measured use. Adoption reports may miss some use. A change affecting the whole market may be hard to see in a comparison within that market.

Those are reasons to interpret the design, not excuses to select whichever result we prefer. The Danish findings restrain a sweeping claim that adoption must already have produced large average labor-market effects everywhere studied. The U.S. pattern restrains a sweeping claim that there is nothing to worry about for younger workers in exposed occupations.

The sensible response is to continue measuring the entry route carefully. We should be especially interested in new hires, the content of junior work, time spent teaching and what happens to people who would previously have entered those roles.

A debate that asks only whether AI is good or bad for jobs is too blunt to do that work. We need to know which decision changed, for whom, over what period, and compared with what.

One percentage can contain a different population

The Bureau of Labor Statistics reported that 53.8 percent of people aged 16–24 were employed in July 2026. The youth unemployment rate was 9.1 percent, down from 10.8 percent a year earlier. These are U.S. summer figures, not seasonally adjusted. They include students and people who did not attend college, and they do not identify an AI effect.³

At first glance, a reader might try to subtract the unemployment rate from one hundred and wonder why the employment share is so much lower. The reason is the denominator: the group used to calculate each percentage.

The employment-to-population share counts employed people as a share of everyone in the age group. The unemployment rate counts unemployed people as a share of the labor force, which includes employed and unemployed people. The unemployed generally have no job, are available and have actively sought work; people on temporary layoff have a separate route into the count. Someone outside the labor force is in the first denominator and outside the second.

A student who is not looking for work is different from a person actively looking and unable to find it. The statistics need a way to keep those situations separate. They do not, by themselves, tell us whether either person is content with their circumstances.

Youth rates use different denominators

Measure	Value or meaning
Scope	United States, July 2026; ages 16–24, civilian noninstitutional population. Not seasonally adjusted.
Population	39.647 million people = 100% of the first bar.
Employed	21.320 million; 53.8% of population.
Unemployed	2.131 million; 5.4% of population.
Outside labor force	16.196 million; 40.9% of population.
Labor force	23.451 million people = 100% of the second bar.
Employed share	90.9% of labor force.
Unemployment rate	2.131 ÷ 23.451 × 100 = 9.1%.
Reading the bars	Both bars run from 0% to 100%, each with its own denominator. Segments run left to right in the listed order. Equal bar lengths do not mean equal numbers of people.
Limits	Survey estimates include students and non-graduates, not just new entrants. These counts do not identify AI effects. Rounded population shares total 100.1%; geometry uses unrounded shares.
Axis ticks	0, 50, 100

Source: BLS data; Definitions

Figure 1. The denominator changes what a labor-market percentage means.

The figure reconstructs the measures from the official published counts. It does not turn the data into a new explanation of why young people are or are not working. That distinction is useful: a calculation can be original and reproducible while making only a modest contribution. Here its job is to prevent us from comparing rates that mean different things.

A broad youth market can improve while a particular route into skilled work becomes harder. The broad figure does not erase the narrower problem. Nor does the narrower problem justify saying that all young people are losing work.

This is why we should hesitate before sharing a dramatic chart whose age group, time period or outcome is hard to find. Those details are not clutter around the result. They define the result.

A degree and a fitting job are different outcomes

The New York Fed’s second-quarter 2026 estimates put recent college-graduate unemployment near 5.6 percent and underemployment near 42 percent. Its recent-graduate group is aged 22–27. Underemployment here classifies employed graduates by whether their occupation ordinarily requires a degree; it does not directly measure low pay, unhappiness, unused talent or education’s value.⁴

The distinction matters to a person trying to interpret their own start. Working in a role that does not usually require a degree may provide income and valuable experience. It may be a chosen direction. It may also be a frustrating stop while a person searches for work closer to what they studied. The classification does not decide which story applies.

A useful conversation asks about the actual job. What are you learning? What does it pay? Is there a route forward? Are you staying because the work suits you, because it is the best available option or because every other route has remained closed?

We should ask training providers similar questions. If a course advertises career opportunity, what happens to participants after it ends? A placement count can conceal major differences in hours, pay, duration and relevance. A graduate working one short contract and a graduate entering a stable development role should not be treated as interchangeable successes simply because both have some employment.

The goal is to find out whether the route helps people build a life and increasing ability. The degree label is one piece of that account.

It would also be a mistake to make graduates stand for all beginners. An entry system that works only for people who have completed a long course of study leaves other routes unexplained. We need to examine apprenticeships, school-leavers, adult entrants and people whose earlier experience is difficult for an employer to read.

What would change our minds?

A strong opinion becomes more useful when we say what evidence could revise it.

My concern is that some organizations may improve current output while weakening the opportunities that produce future experienced workers. Evidence of paid entry hiring, protected teaching time and good progression in those organizations would reduce that concern. Evidence that they have eliminated beginner roles, increased experience requirements and left no replacement route would strengthen it.

At the economy level, we would want longer observation, more than one data source and ways to separate AI from other changes. Interest rates, demand, sector shifts and earlier hiring cycles can all be relevant hypotheses. A serious explanation should test alternatives where the data allow it.

At the learner level, we would want more than a count of completed tasks. Can the person explain their choices, detect a problem and handle a changed case? At the program level, we would want to follow those who leave as well as those who finish. A scheme can look successful by reporting only its best survivors.

None of this requires an applicant to become a labor economist before sending a résumé. It is a standard for the people making large claims and selling large solutions. If someone says a costly credential will protect your future, they owe you evidence about the path it opens. If a company says it has solved training with a new tool, ask what the learner can do afterward and how the company knows.

Clear questions are a form of protection against expensive confidence.

Use the evidence at the right scale

There is a temptation to read a labor-market report as a personal forecast. You see a poor result for a group you belong to and imagine the line extending through the rest of your life.

The report cannot do that. It describes a group, over a period, under conditions that can change. It may reveal a real disadvantage. It cannot measure every connection you will make, every skill you will acquire or every institution that might offer a chance.

There is an opposite temptation too: to tell a successful personal story as if it disproves a group disadvantage. One person finds work quickly, so everyone else must be doing something wrong. That is no better. A route can remain open for some while becoming much harder overall.

We need room for both the person and the pattern. The person deserves practical help. The pattern deserves a public response when it shows a problem beyond individual control.

So read the numbers, but keep them in their proper place. A hiring warning should lead to better questions about entry. A successful program should lead to questions about what made it work and whether those conditions can be built elsewhere. An encouraging average should not close the conversation with people whose route is still blocked.

The next task is closer to the ground: deciding what a beginner needs to learn, and how a workplace can tell that the learning is happening. The door matters. So does what happens after it opens.

Sources and limits