Is AI Taking Entry-Level Jobs? Only Where It Substitutes for You
Two findings that look like they cancel out.
Stanford economists tracking real payroll records through June 2026 find workers aged 22 to 25 in the most AI-exposed jobs employed 19% below where they would otherwise be. An analysis of 1.3 billion job postings finds roles asking for AI skills advertising about $18,000 more.
Fewer doors, better pay behind the ones still open. Both numbers are sound. Here is why they do not cancel out, and what to do about it.
The damage is real and narrower than the headlines
The Stanford paper is called Canaries in the Coal Mine?, it was revised this month, and its first finding is the one that rarely gets quoted: there is no evidence of widespread, economy-wide job displacement. The labor market as a whole has not fallen over.
What has happened is specific. The 19% gap applies to young workers in Jobs whose day-to-day tasks overlap heavily with what current AI systems can do. Software development, customer support and routine writing and analysis score high; anything built on physical presence, in-person judgment or accountability someone has to personally carry scores low. Exposure measures task overlap, not whether your employer has actually adopted anything. , and experienced workers in those same occupations show no comparable gap. The divergence has widened steadily since it was first spotted in August 2025, so this is a trend rather than a blip.
That 19% is routinely misread, so be precise about it. It does not mean 19% of these jobs disappeared. It is a A comparison against what would have happened otherwise, rather than against the past. Employment for this group sits 19% below where it would be if it had kept pace with comparable workers in less-exposed jobs. The gap can open up while the absolute headcount is flat or even slightly rising, which is what makes it easy to miss and easy to overstate. gap: hiring of 22 to 25 year olds in those roles is running 19% below the path it would otherwise have followed.
It shows up as a hiring freeze, not a layoff wave
The mechanism matters more than the number, and it is the most practically useful thing in the paper. The change is happening primarily through reduced hiring rather than increased separations. Firms are not clearing out their juniors. They are quietly not replacing them.
Another finding completes the picture: the adjustment is happening through employment rather than base compensation. Employers are not cutting the pay of the people doing this work. They are cutting how many of those people they take on.
That tells you where the risk actually sits. If you already hold one of these jobs, the thing to watch is not your salary or a redundancy round. It is whether the person who left last quarter got backfilled. If you are trying to get in, the door is narrower than it was, and it narrowed without any announcement.
Substitute or complement: the difference that decides your job
This is the finding that matters most. Declines concentrate in occupations where AI substitutes for human tasks. Where AI complements the worker, employment is flat or rising, and especially so for experienced people.
Same technology, opposite outcomes, and the difference is not the industry or the job title. It is the shape of the work.
The same technology, two outcomes
Substitute or complement
What decides it is not your industry or your job title. It is the shape of the work.
Substitute
AI finishes the task.
A good result means nobody else needed to weigh in.
19%
below where entry-level employment would otherwise be
By hiring less, not by layoffs
Complement
AI drafts. A person decides.
It still needs someone to check it and answer for it.
Flat or rising
employment in the same occupations, most of all for experienced workers
The test: when AI does your main task well, is the task finished, or is it a first draft?
A rough test you can apply to your own week. Take the task that occupies most of it and ask what happens when AI does it well:
- If a good result means the task is finished and nobody needed to weigh in, that task is being substituted.
- If a good result is a draft, an option or a first pass that still needs someone to decide, check it and answer for it, that task is being complemented.
Most jobs contain both. The question is which one your week is mostly made of, and which direction you are drifting. This is the individual version of the exercise managers are running on whole roles in AI Role Redesign, and the task-level arithmetic behind it is in How Much of Your Job Can AI Actually Do?
Which is why the pay premium isn’t a contradiction
Now the 28% makes sense. Lightcast analyzed over 1.3 billion job postings and found that ones listing AI skills advertise about $18,000 more. The detail that matters for most readers of this site: 51% of those postings sit outside IT and computer science, and demand for generative-AI skills in non-technical roles has grown by around 800% since late 2022. This is not a programming story.
So the two numbers describe the two sides of the same split. Roles where AI substitutes are quietly not being refilled. Roles where AI complements are being advertised at a premium, because someone who can direct the tool and stand behind the output is worth more than someone who cannot.
Two honest caveats. The Lightcast figure is a correlation, not a raise: postings that mention AI skills also skew senior, better funded and better paid for reasons that have nothing to do with AI, and nobody has shown that adding the skill to your CV causes the premium. And that report is from July 2025, which is old for this subject.
What to actually do about it
- Sort your main tasks into substituted and complemented. Ten minutes with the test above tells you more about your position than any headline about AI and jobs.
- Move toward the part someone has to answer for. Judgment, relationships, deciding what the answer should be and being accountable when it is wrong. That is the complemented half, and it is where the employment numbers are flat or rising.
- Name the skill explicitly. Half of the AI-skill postings are outside IT, so “I use AI for X in my role” is a claim a non-technical CV can credibly make. How to Learn AI at Work When Nobody’s Training You covers building something worth naming.
- If you are early in your career, take the hiring signal seriously. The exposed occupations are the ones not opening entry-level doors. The complemented ones still are, and that should influence where you apply.
The takeaway
AI is not taking jobs in general. It is thinning out the entry level of the occupations where it can finish the task on its own, and it is doing it by quietly hiring fewer people rather than by letting anyone go.
Work out whether AI finishes your main task or just speeds it up. If it finishes it, move toward the part that needs a person to decide and be answerable. That is the same split the pay premium is measuring. It is also the one thing in all of this you can actually act on.
Sources: the 19% figure, the hiring-versus-separations mechanism and the substitute-versus-complement split are from Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (opens in a new tab) (Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab), using ADP administrative payroll data covering millions of US workers through June 2026, revised 12 August 2026. The pay figures are from Beyond the Buzz (opens in a new tab) (Lightcast, July 2025), an analysis of more than 1.3 billion job postings. Lightcast sells labor-market data, and the report is now a year old. The substituted-or-complemented test is mine, written to make the paper’s distinction usable on a Monday; the research establishes that the split exists and predicts employment, not that any individual can reclassify their own job by choosing to.
Related: How Much of Your Job Can AI Actually Do? About a Fifth puts a number on the task overlap this post depends on. AI Anxiety at Work: You’re Stressed and More Productive. Both Are True. is the companion for how this lands on the people living through it.