Payroll counts put the developer employment drop inside the 22-to-25 age band

Payroll tells you who got paid. A job posting tells you what someone typed into a req. The dev.to write-up leads with that distinction, and it holds up: a role advertised as junior and then filled by a senior still registers in the posting count as a junior opening ^[3]. Nothing stops a company from advertising at one level and hiring at another, and nothing in the posting count notices. The Stanford Digital Economy Lab analysis in 2025 used payroll instead, and found employment for software developers aged 22 to 25 down by nearly 20% from its late-2022 peak ^[1]. The headlines the post is arguing with put the loss at two-thirds of entry-level programming jobs ^[10], about three times the payroll figure ^[13].

For that 20% to describe your own bench, age has to stand in for experience level, since the study's cut is an age band and a junior population is a tenure band ^[1]. Your roles need to sit inside whatever the study counted as highly AI-exposed; the post does not say how the study classified exposure, what the sample was, or where the data came from ^[12]. And late 2022 would have to be your peak too, because the industry over-hired during the pandemic boom and then spent years correcting ^[4].

The comparison sitting next to the drop is harder to dismiss. Developers aged 30 and older, in the same AI-exposed roles over the same window, did not decline; their employment grew ^[2]. Both cohorts sat inside the same interest-rate environment and the same post-boom correction ^[4]. A correction that lands on the most recent hires first would also show up as an age gap, so the split narrows the candidate causes without settling them. The author is direct about that: he wrote that anyone selling a clean story in which AI did it is overselling ^[8].

Adoption cuts against the simplest version of the story as well. The 2025 Stack Overflow survey has about 84% of developers reporting they use AI tools ^[5], which leaves 16% reporting they do not ^[6]. At that level, an explanation resting on which cohort has the tool has very little room to work.

So the payroll data supports a location, and the cause stays open. The post's prescription is aimed at the individual developer, who is told to stop being interchangeable with what AI does well and to build judgment, context and the trust of the team ^[11]. The work it identifies as exposed is the well-defined, low-context task, the kind a model now does cheaply ^[9]. Who gets handed that work is settled higher up, by whoever assigns the tickets. On the size of the effect, the author wrote that "the pressure is concentrated at the entry point, not across the field" ^[7].