Quick answer: On August 12, 2026, Stanford's Digital Economy Lab published a revised version of its study on AI and employment. Workers aged 22 to 25 in the most AI exposed occupations now hold about 19 percent fewer jobs than they would if they had kept pace with people the same age in less exposed work. That gap was 15 percent in July 2025. Workers over 30 in those same occupations show no comparable gap, and in roles where AI assists rather than replaces, their employment is flat or rising. Jobs are not vanishing across the economy. The bottom rung of the ladder is.
A chart about this has been circulating all week, usually with a headline claiming AI is wiping out jobs. The underlying research says something narrower than that, and the narrow version is far more useful if you actually run a company and have to decide who to hire next quarter.
Here is what the data says, what it does not say, and what to do about it.
What the study found
The project is called Canaries in the Coal Mine, run by a team at Stanford including Erik Brynjolfsson. It works from payroll records covering millions of workers rather than survey responses, which is why it gets taken seriously. The August 2026 revision added a much larger dataset than the first version, and the picture got sharper rather than softer.
Three findings matter:
- The gap is widening. Employment for 22 to 25 year olds in highly AI exposed occupations sat 15 percent below the comparison group in July 2025. By June 2026 it was 19 percent. It is not levelling off.
- It is concentrated, not general. Same age group, less exposed occupations, no such decline. The researchers are explicit that they do not see widespread displacement across the economy.
- Experience is protective. Workers in their thirties and beyond in the very same occupations show no comparable gap. Where AI complements the work instead of automating it, their employment is flat or growing.
The line that actually separates safe work from exposed work
This is the part worth reading twice, because it is more useful than any list of at risk job titles.
The declines showed up in work built on what the researchers call codified knowledge. That means formal, standardized, documented knowledge, the kind you can learn from a textbook, a certification or a well written manual. Work that depends on tacit knowledge, the sort you only pick up by doing the job and getting it wrong a few times, held steady or grew.
Put plainly: if the knowledge a role runs on has been written down somewhere public, a model has already read it. If the knowledge lives in someone's head because they learned it from six years of awkward client meetings, it has not.
That explains a pattern a lot of managers have noticed without being able to name it. The first year analyst who produces a standard report is under real pressure. The person who knows which number the client will argue about is not.
Why this is a problem you inherit later
The obvious business response is to stop hiring juniors. It looks efficient. One senior person plus decent tooling does what three people used to do, and the numbers work this quarter.
The trouble is where seniors come from. Nobody has ever produced a senior engineer, a senior accountant or a senior consultant except by putting a junior through several years of real work. If the industry stops running that pipeline, the shortage does not show up now. It shows up in about five years, when the people you want to hire do not exist and the ones who do cost triple.
Some of that cost lands on you individually and some of it lands on everyone at once. Either way, a hiring freeze at the bottom is a decision to buy expensive talent later instead of building cheaper talent now.
What to do if you are hiring
Hire juniors for judgment, not for volume
The old junior job was throughput. Produce the draft, run the numbers, clear the queue. That work is now cheap. The junior worth hiring is the one who can look at what a model produced and tell you it is wrong, and say why. Interview for that directly. Hand a candidate a plausible but flawed AI output and ask them to review it. You will learn more in ten minutes than from any take home task.
Redesign the first six months
Most onboarding plans still assume the new person learns by grinding through simple tasks for half a year. If AI does those tasks now, the learning has to come from somewhere else. Put juniors in front of clients earlier. Have them sit in on the messy decisions rather than receiving the tidy outcome. Give them ownership of something small and real instead of pieces of something large and abstract.
Write down what is currently in people's heads
There is an irony here that is worth using. Tacit knowledge is what protects your senior staff, and it is also your single biggest operational risk, because it walks out the door when they do. The teams that come out of this well are the ones documenting how decisions actually get made, not just what the process diagram says.
Check which of your own roles are codified
Go through your team's work and mark which parts run on documented, standardized knowledge. That is your exposure map, and it is more honest than any generic list of jobs AI will replace. We covered the practical side of pointing AI at that kind of routine work in our write up on building a junior ops assistant in two hours.
What the researchers themselves are careful about
Worth knowing, because the people posting this chart usually leave it out. The Stanford team notes that the gaps shrink once you adjust for education levels. Some of these trends existed before generative AI arrived. And they say plainly that one study is not definitive evidence. Interest rates, hiring cycles and the general state of the graduate market all sit in the same data.
So treat this as a strong early signal rather than a verdict. That is how the authors frame it, and it is the honest way to use it.
The short version for a small company
- Do not read this as proof that jobs are disappearing. Total employment is holding up.
- Do read it as proof that entry level work built on documented knowledge is getting automated fast.
- Keep hiring juniors, but hire them for review and judgment instead of output.
- Rebuild onboarding so people learn from real decisions rather than from routine tasks that no longer exist.
- Document the tacit knowledge your seniors carry, before it leaves with them.
The uncomfortable part is that waiting several years to slowly become senior is no longer a plan that works on its own. That is true for the people you hire and it is true for how you develop them. If you want the broader view of where automation is landing across business functions, our overview of current trends in AI business automation covers the ground.
Frequently asked questions
Is AI actually destroying jobs?
Not across the economy, according to this research. The Stanford team states clearly that they find no widespread displacement. What they find is a sharp, concentrated decline in entry level employment within occupations most exposed to AI, while employment for experienced workers in those same occupations holds up or grows.
What does the 19 percent figure actually mean?
It is a gap, not a layoff count. Employment among 22 to 25 year olds in highly AI exposed occupations is about 19 percent below where it would be if it had grown at the same rate as employment for people the same age in less exposed occupations. It measures a shortfall against a comparison group, not jobs cut.
Which kinds of roles are most affected?
Roles that run on codified knowledge, meaning formal and documented knowledge that can be taught through education and manuals. Roles that depend on tacit knowledge built up through experience and practice have held steady or grown.
Should we stop hiring junior staff?
We would not. The junior role needs redesigning, not deleting. If nobody trains juniors, the supply of experienced people dries up in a few years and gets expensive for everyone. Hire fewer if you must, but hire them into work that builds judgment quickly.
How reliable is this study?
It uses payroll records for millions of workers rather than surveys, which makes it stronger than most reporting on this topic. The authors still caution that the gaps narrow when education is accounted for, that some trends predate generative AI, and that a single study is not definitive.
What should a young professional do with this information?
Build the parts of the job that are not written down anywhere. Client judgment, scoping, knowing which problem is worth solving, being able to check an AI's work rather than only produce work. Those were once considered senior skills and are now closer to the entry ticket.
Working with us
Buinsoft is a Prague based AI and software consultancy. A good part of our work is helping small and medium companies work out which parts of their operation are routine enough to automate sensibly, and which parts depend on people who should be doing something more valuable with their week.
You can read more about our AI integration consultancy, email us at info@buinsoft.com, or get in touch through our contact page.
