Entry-level software engineering postings in the United States dropped 67% between 2023 and 2024, according to research from Stanford's Digital Economy Lab working with ADP payroll data. Google and Meta are hiring roughly 50% fewer new graduates than they did in 2021. Salesforce publicly announced it would halt junior hiring for 2025. Those are the numbers everyone quotes, and they are real.
The conclusion people draw from them, that AI has eliminated the junior developer, is the part I keep getting stuck on. The data is messier than the headline and I think the messy version is more useful if you are actually trying to get hired right now.
What the research actually found
The most careful finding in the Stanford work is narrower than the 67% figure suggests. When companies adopt generative AI tools, junior employment in affected roles drops 9% to 10% within six quarters. That is a meaningful effect and it is specifically attributable to AI adoption. It is also nowhere near 67%.
The larger collapse in postings overlaps with the end of zero interest rate policy, the correction after the 2021 and 2022 hiring bubble, and a broad round of cost cutting across tech. Listings are down roughly 70% from their 2022 peak but only about 35% from pre-2020 levels. That second number matters. It means a large chunk of what vanished was the bubble, not the baseline.
There is also contradicting evidence worth sitting with rather than explaining away. Seth Robinson, VP of industry research at CompTIA, has reported dramatic increases in software development job postings in 2026. The US Bureau of Labor Statistics still projects 15% growth for software developers from 2024 to 2034, roughly five times the average across all occupations. I do not know how to fully reconcile those with the Stanford numbers, and I am suspicious of anyone who claims they do.
The part that is genuinely about AI
Here is where I think the AI explanation holds up. The tasks that used to constitute a junior developer's first year are precisely the tasks these tools are good at: boilerplate, simple bug fixes, unit test scaffolding, basic refactoring, documentation.
That was never busywork. It was the on-ramp. You learned a codebase by writing its tests. You learned the domain by fixing small bugs in it. You learned what good code looked like by refactoring bad code under review. Those tasks taught you the thing that made you a mid-level engineer two years later.
The problem is not that AI does those tasks. It is that nobody has replaced the learning they used to provide. A junior in 2026 is expected to arrive already fluent with AI coding tools, capable of shipping complete features rather than isolated functions, and producing at a level that previously took two to three years to reach. The industry removed the training ground and kept the expectation that people would somehow arrive trained.
What I would do differently if I were starting now
I am not going to pretend I have a clean answer. But a few things follow from the above.
Get very good at the part AI is worst at, which is judgment about systems that already exist. Generating a function from a clear spec is solved. Figuring out why this specific production system behaves badly under this specific load, when the cause is spread across three services and a cache, is not close to solved. That work requires holding a real mental model of something messy, and models are still bad at it.
Learn to read code far more than you learn to write it. This inverted at some point and I do not think the advice caught up. If a tool can produce 200 lines in eight seconds, the bottleneck is whether you can tell that lines 40 through 60 are subtly wrong. That skill comes from reading a lot of real code, including bad code, and it is the thing that makes someone useful on day one now.
Ship things end to end. The bar moved from "can write a function" to "can take a vague requirement and produce a working, deployed, tested feature." That is a much bigger scope, and side projects are one of the few places you can practice all of it without permission.
Be genuinely fluent with the tools rather than either dismissive or credulous about them. I wrote a longer comparison of Copilot, Cursor, and Claude Code and the honest summary is that they each fit different tasks and the fluency is in knowing which. Being the person on a team who knows when not to reach for the agent is worth as much as knowing how to prompt it.
The thing that actually worries me
Not that juniors get replaced. That the pipeline breaks and nobody notices for five years. Mid-level engineers come from juniors. Seniors come from mid-levels. If the industry stops hiring at the bottom because AI covers those tasks, the supply of people who can review AI output competently dries up on a delay long enough that the causation gets hard to see. By the time that bites, the people who would have been fixing it went into other fields in 2025.
Dice's careers coverage put it in a phrase I keep coming back to: the new junior developer is harder to land and harder to replace. Both halves are true. The role got more demanding and the people who clear the higher bar are more valuable than juniors used to be. That is not much comfort if you are currently sending applications into the void, and I do not want to dress it up as one.
Where I land
I do not think this is the end of entry-level work in software. I think it is a genuinely bad few years to be entering the field, caused by a market correction and an AI shift arriving at the same time, and the two are getting collapsed into one story because that story is simpler. If you are in it right now, the market is worse than the people telling you to just build a portfolio seem to understand. It is also not the extinction event the more dramatic posts describe. Both of those can be true, and I would rather say that than pick whichever one is more shareable.
Sources: Stanford Digital Economy Lab, US Bureau of Labor Statistics occupational outlook.