I have spent twenty-eight years building things on computers, and I have watched the “this is the end of developers” headline come around more than once. So when the latest round of articles proposing the end of developers because of AI started landing, I read them, and continue to read them, with interest rather than panic. They are worth taking seriously. Their conclusions are also wrong.
In 2025, Stanford’s Digital Economy Lab looked at payroll data covering millions of workers and found something specific and real. Employment for the youngest software developers, ages 22 to 25, fell about 20 percent from its peak, while older developers in the same field held steady or grew.
Entry-level postings shrank. New-grad unemployment still appears to sit above the national rate.
If you are about to graduate, or you just did, that is a frightening picture, and I am not going to tell you it is imaginary.
But the headline that came out of those numbers, that AI is erasing the entry-level job, mistakes what actually happened in my experience. The jobs did not disappear. They changed. And the same thing is happening not just to software developers, yet also to junior lawyers, junior analysts, and a lot of other people walking in the front door of a profession. The shape of the change is the same everywhere, so stick with me even if you do not write code.
Here is the lens I want you to carry through the rest of this post: AI is a multiplier of judgment you already have.
Hold that thought.
What actually moved
The entry-level developer job used to be “write the code” (or write a brief, or analyze a business). You got hired to take a well-specified ticket and turn it into working code, and you got better by doing a lot of it. Same thing happened to software developers in the 60’s, punch assembly into cards, input, review, repeat.
That first rung is gone, or going. AI writes that code now (or brief, or analysis). What replaced it is a different first job: judge and direct the code (output). Can you read what the machine produced and tell whether it is right? Can you spot the design that will collapse in six months? Can you take a vague problem, break it into pieces, hand those pieces to AI, and verify what comes back? AI is certainly getting better, but it has an inherent flaw in reasoning.
That is the job now. It is not harder or easier in the abstract. It is different in kind. And it lands hardest on the person who has not had time to build the judgment it requires.
The multiplier
Here is why seniors are pulling ahead while juniors struggle, and it is not that seniors are smarter.
A multiplier acts on whatever you bring it. A developer with twenty years of judgment, multiplied by AI, becomes dramatically faster. A developer with no judgment yet, multiplied by the same AI, is still close to where they started, and sometimes worse, because now they can produce broken work at high speed without knowing it is broken. I’m seeing this with non-developers that think vibe coding will make them rich by rebuilding Salesforce over the weekend and rushing to get paid customers. Good luck.
The data backs this up in an uncomfortable way. Sonar’s 2026 developer survey found that 96 percent of developers do not fully trust that AI-generated code is correct, yet only 48 percent always check it before committing. While this study is a few months old, even with the latest frontier models, that gap, between knowing the code is suspect and actually verifying it, is the whole danger. The same survey found that junior developers report the biggest productivity gains from AI and also find reviewing AI code harder than their senior colleagues do. They get the most speed and have the least ability to check the output.
That is the multiplier working exactly as you would expect. Zero times a large number is still zero.
We need to be honest about the evidence
I am not going to cherry-pick. The story is messier than either the doom crowd or the boosters want it to be.
The layoffs are real. Tech cut around 150,000 jobs in the first half of 2026, and plenty of them were pinned on AI. That is not a phantom, and if you lost a role or could not land a first one, you already know it.
Now look at who is actually getting hired, because the layoff headline hides it. SignalFire tracked careers across tens of millions of companies and found engineering was the most resilient job function in tech. While overall hiring at big tech fell 25% from 2019 levels, engineering roles fell only 11%, and engineers grew to 55% of all new hires in 2025 at the twelve largest tech companies, up from 46% in 2019.
Early-stage startups hired 7% more engineers in 2025 than they did in 2019. Set that next to the Stanford number and the picture sharpens. The function is thriving. It is the youngest, least-experienced slice of it that is getting squeezed. The resilience is concentrated exactly where the judgment is.
But watch what happened next. Companies are quietly rehiring. Forrester found that 55% of employers regret AI-driven layoffs! Gartner projects that half of the firms that cut jobs for AI reasons will rehire for similar roles!
Nearly a third of hiring managers who eliminated roles for automation have already rehired humans for those exact positions, because the models turned out not to understand the company, the customer, or the context. They cut judgment they could not actually do without, and they are buying it back. This trend will continue, frontier models will have limitations.
Two honest caveats
First, the rehiring does not cancel out the layoffs. It is not offsetting the cuts, and it skews toward experienced people, not new grads.
Second, not everyone agrees AI is even the cause. The New York Fed found that the decline in postings for AI-exposed jobs started before ChatGPT existed, and that there is no clear gap between junior and senior hiring within those jobs, which makes it hard to blame AI alone for the entry-level slowdown.
Some of what got labeled an “AI layoff” had little to do with AI. A lot of these cuts were the hangover from pandemic-era over-hiring, when cheap money paid for roles that higher interest rates and pressure on margins no longer justify. “We are leaner because of AI” is a better story to tell investors than “we hired too many people when money was free,” so it’s a safe bet the label got stuck on cuts that are really about the balance sheet.
You can see the same pattern in a place that has nothing to do with software. Junior lawyers used to earn their stripes doing deep legal research. AI does the first pass now. But courts are sanctioning attorneys for filing briefs full of cases that do not exist, as they should, because of citations the AI invented and nobody checked. Sure, there are companies attempting to fix this problem, it will get better, though it will not make the AI problem go away.
A running database of these incidents maintained by Damien Charlotin has logged more than 1,600 of them. I verified a sample myself rather than trust the list on faith, which is an underlying point I’m making. Anecdotally (my judgment), it looks right as well. Anyhow, in several of those cases, the bad filing traced back to a junior or an unlicensed clerk who trusted the output and did not verify it. Isn’t that just the developer’s “looks correct but isn’t” problem, yet in a courtroom, with a fine attached!
The job that survives in law is the same one that survives in code: the judgment to know whether the machine’s confident answer is actually true. Beyond just the “fact-check”.
A word to the people doing the hiring
If you run a team, the knee-jerk move was to stop hiring juniors and let AI cover the bottom. I think that is a mistake you will pay for.
You cannot keep harvesting senior judgment if you stop growing any. Every senior you rely on was once a junior somebody invested in. Freeze that pipeline and you are eating your seed corn. The rehiring wave is the early bill coming due. There is a smarter move than cutting, and the strongest teams are already making it.
When engineers became more productive with AI, the work did not run out. It expanded! To fill the new capacity! This is a pattern economists call the Jevons paradox. Nvidia’s CEO, Jensen Huang, says his engineers are busier than ever now that they run agents, constantly pushed toward the next idea.
Keep your people, multiply them, and you outrun the competitor who banked the savings and shrank. The firms that win the next decade will be the ones that pair juniors with AI and grow judgment faster, not the ones still pretending they can run on seniors forever.
I also keep seeing experienced people take early exit ramps out of the big companies, and a lot of hard-won judgment is leaving the building with them. That is a real loss, and it is also an opening.
If you are the junior reading this, the person who can walk in and supply real judgment to a fleet of agents from day one has room to rise that the old ladder never offered. This is why your moment is coming, as long as you build the judgment to take it. The correction is already starting.
What is actually durable, and the bet I am asking you to make
So where do you aim? My answer is architecture and verification, the judgment about how systems fit together and whether they hold. Yes, this is a bet, AI cheerleaders seem to say AI will take it all. Ok, I’ll tell you why I make this bet anyway.
AI got real good, real fast, at things with a clear right answer. Code either compiles and passes the test or it does not. That clean feedback is exactly what let the models improve so quickly.
Architecture, judgement. They have no unit tests.
There is no instant grader for “was this the right boundary between services.” Further, you often do not find out you were wrong for a year or more.
That is the structural reason the gap is likely to stay open, not just that it is hard today. Yes, frontier models are impressive at finding bugs in system. Will you trust their answer, their judgement, with your life or livelihood?
Sonar’s survey shows it from the practitioner side too: developers rate AI highest at boilerplate and working within existing context, and weakest at the nuanced work of refactoring and maintaining mission-critical systems without close supervision. I’m seeing the same output in other spaces I work. Boilerplate docs = great. Actual reasoned content = I need to ensure the output is correct or my client will (and should) fire me.
Now here is why the bet is safe even if I am wrong about the timeline. If the gap stays open, you do the architecture yourself and AI cannot take it from you. If the gap closes and AI gets great at architecture, then the person who wins is the one who can direct and verify that work, and you can only do that if you understand it.
Either way, understanding the system pays off. You are not betting on AI staying bad. You are betting that understanding never stops mattering, which is a much safer bet.
Care about the output, but know which corner you are cutting
In college, I studied all things computers, i.e. software development and electrical engineering. I took little interest in circuits. What I wanted was to make things, to bend a computer to my will and have something real come out the other side. Software was the medium that let me do it. I never cared that much about the language. I cared about the output.
If that is you, AI is the best gift you have ever been handed. I can build things now that were never economical for one person to build. That is genuinely thrilling, and you can feel it. I sure do.
But “I care about the output, not the details” is also exactly what the person typing “build me an enterprise app” into a chat box says. Same words, opposite outcome. The difference is what you mean by output.
If you mean the real thing, the one that works and keeps working when you have actual customers, then caring about output drags you straight into the craft, because you cannot get there without it. If you mean the demo that looks fine on the front end, you will skip the craft and ship a beautiful thing that falls apart the moment it meets the real world.
We cut corners all through history, and most of the time it is fine. Very few of us can ride a horse (at least safely, with proficiency) anymore, and we are doing okay. The trap is cutting the corner that builds your judgment and telling yourself it was just busywork.
The junior who lets AI write code they cannot read, and the associate who files a brief they did not check, both cut that corner. Cut the mechanical stuff freely. Never cut the corner that teaches you to tell good from garbage.
The playbook
Here is where I would put your energy, in order. I will name the underlying skill for each, so you can map it to your own field if you are not a developer.
Learn the fundamentals by building real things yourself. I still believe in writing actual programs in a real language. C++/C#/pick your deeper language here to understand what the machine is really doing, Python/JavaScript/TypeScript to get things done. Don’t nitpick me here, yes, there are many great languages, I work in many different ones near each day. The point is not fluency in every language. It is that, having built things by hand, you can read almost any code and reason about what it is doing. The underlying skill: you cannot verify what you do not understand.
Get genuinely good at verification and QA. This is your way in the door. You are not going to become a great architect in six months, but you can become excellent at testing, reviewing, and breaking AI output in that time, and that skill is worth money the day you have it. The underlying skill: telling real output from plausible garbage.
Build toward architecture / reasoning. This is the long game and the deep moat. It is the judgment that turns you from someone who completes tickets into someone who decides how the thing should be built. This is where you stop being a junior.
Learn to orchestrate. Once you can build, verify, and reason about design, you can run a team of agents instead of doing every keystroke yourself. That is a skill of decomposition and supervision, and it sits on top of everything else, not instead of it.
Through all of it, the way you get hired is by showing, not claiming. Build real things now, while you are job hunting, and put them where people can see them. A project that survived contact with real users says more than any certificate. Spend your learning time on things you can point to.
Build your own apprenticeship
Here is the hard part, said plainly. The paid junior years that used to teach all of this, the years where a company paid you to learn on the job, are shrinking. So you are going to have to fund the apprenticeship yourself, with the same AI that raised the bar in the first place.
That is not as bleak as it sounds. The tool that made the entry harder is also the fastest learning accelerator anyone has ever had. You can build, break, and rebuild more systems in a year, possibly in a month, than I could in five years when I was in school. Use it that way. Build something real, break it on purpose, figure out why it broke, and do it again. That is your apprenticeship, and you are the one paying for it, in effort.
There is a catch, because pretending otherwise would not help you. This rewards the self-directed. The person who reads the situation and starts building without being told gets a head start the person waiting for permission does not. That is not entirely fair, but it is true, and knowing it is true is itself an advantage.
The door did not close. It moved.
So no, AI did not erase the entry-level job. It just changed. The strange, hopeful part is that the same tool that moved it is the one that lets you reach it.
Go build.
Speak Your Mind