AI replaces junior tasks, not junior people

Every few weeks someone tells me AI will make junior developers unnecessary. I'm early in my own career, so I've thought about this more than most. My conclusion: AI replaces junior tasks, not junior people. What changes is what a junior is paid for.

Is junior hiring really falling?

Yes. I won't pretend otherwise.

• US, by age: Stanford's "Canaries in the Coal Mine" study uses payroll data from ADP. In its August 2026 update, employment of 22- to 25-year-olds in the jobs most exposed to AI was 19% below where it would have been had it kept pace with less-exposed peers. Experienced workers show no such gap. The authors say the gap comes mainly from fewer young people being hired, not from people being let go.
• Big tech: SignalFire's 2025 talent report found new graduates made up just 7% of hires at the 15 largest tech companies, with new-grad hiring down more than 50% compared to 2019.
• Germany: Indeed's Hiring Lab reported in September 2026 that entry-level job postings (titles with Junior, Trainee or Graduate) were 46% below 2019 levels in July 2026. In software development, there are now about three entry-level postings for every ten in 2019.

So the objection to my thesis is real, and I'll come back to it. But look closer at what is disappearing.

What does AI actually replace?

The Stanford study has a detail that matters more than its headline: the declines are concentrated in jobs where AI substitutes for human tasks. Where AI complements people, employment is flat or rising.

Think about what a junior developer traditionally did in their first year: write the boilerplate, build the first draft of a function, set up the tests, write a SQL query someone else designed, fix the small bugs. That is exactly the work AI now does in seconds. In a September 2026 survey by Bitkom, the German digital association, 55% of companies said they expect "classic entry-level tasks" to increasingly disappear.

Note the wording: tasks. The same survey found 85% of companies had not cut IT jobs because of AI. And for Germany, Indeed's own economist said the current data does not show an AI effect yet. The weak economy explains more. The research institute IAB came to a similar conclusion in October 2026, and its one possible warning sign was in manufacturing and non-academic jobs, not among graduates.

Why can't a doctor with ChatGPT do the same work?

This is the part I find most interesting. Today, a medical student or a mechanical engineer can open an AI assistant and get working code. Output is no longer the bottleneck.

But producing code was never the hard part. Ask an AI to "build a data pipeline" and it will build one. Whether it should run in batches or stream continuously, whether raw data should be stored before it's transformed, what happens if the same job runs twice, how you'll notice when it breaks at 3 a.m.: the AI handles all of these well if you ask. To ask, you need to know these questions exist.

That is what studying software gives you, even if you're not the fastest coder in the room. You know the shape of the problem. You can give the right instruction, and you can tell when an answer is wrong.

And answers are wrong more often than they look. In a 2025 test of more than 100 language models by the security company Veracode, about 45% of AI-generated code samples contained security flaws. Someone has to catch those.

The research on AI and productivity fits this. In a large study of customer-support agents, AI raised productivity by 14% on average and by 34% for novices. AI raises the floor of what anyone can produce. But a 2025 study by METR found that experienced open-source developers were 19% slower with AI on their own projects, while believing they had been 20% faster. METR's 2026 follow-up was less clear-cut, but the lesson holds: people are bad at judging their own AI-assisted work. Knowing what good looks like is the skill.

Why does shared language matter more than syntax?

In my first job working with data, I noticed something that surprised me. My colleagues talked almost only in technical terms: pipelines, ETL workflows, microservices, how data should flow from one system to the next. I understood all of it, not because I was experienced, but because I had studied exactly these concepts. I didn't have to look up words in my head. I could follow the discussion and add to it.

That's when I understood what a degree actually buys you: not the ability to type code, but access to the conversation. And the conversation is where the real decisions are made. Which approach is best? What breaks if we do it this way? What will we regret in a year?

Software people have a name for this. Eric Evans calls it a "ubiquitous language" in Domain-Driven Design. Martin Fowler describes it as "a common, rigorous language between developers and users," which needs to be rigorous "since software doesn't cope well with ambiguity."

With AI, this matters more, not less. When the typing gets delegated, the discussion about what to build becomes most of the job. A team where one person can't follow the discussion is slow, however many AI tools it has.

What will a junior role look like?

I think it shifts in three ways.

1. Less writing, more judging. Reviewing AI output, writing the specification before the code, testing whether the result really works.
2. Efficiency is the measure. In the end, a job is about how much good work you deliver in a given time. A junior who uses AI well and knows when to distrust it will outperform one who types everything by hand, and one who trusts everything blindly.
3. Thinking across several dimensions wins. The most useful people understand the business problem, the data and the software at the same time, and can connect them. That's also the reason I work where business, data and software meet.

So if you're starting out: learn the concepts deeply, learn the vocabulary of your field, and let the AI do the typing. Junior positions will change. They won't disappear, because someone still has to know what to ask for.

If you're curious how AI handles different kinds of work, I wrote about where it's strong and where it still struggles in AI speaks code, not clicks.