AI hits junior developers: 7% unemployment and fast food offers

AI eats the junior developer: 7% unemployment and fast food offers

In 2024, over 170,000 students were enrolled in computer science degrees in the US, double the figure from a decade earlier. The promise was simple: learn to code and join an industry offering six-figure salaries. The result, judging by circulating data, is far less epic. Some recent graduates, after sending hundreds or thousands of applications, have only received responses from fast-food chains. The unemployment rate for new computer science graduates hovers around 6%, exceeding 7% for computer engineering. By comparison, fields like biology or art history remain around 3%.

The trigger isn't just reduced hiring. It's that entry-level positions, precisely those absorbing fresh graduates, are the easiest for AI to replace. Amazon, Microsoft, Meta, and Intel have cut thousands of jobs while generative tools perform in seconds what a junior did in hours. The door closes before a recruiter even reads the resume.

Why blame AI instead of interest rates?

Here the disagreement begins. One school of analysis argues AI has nothing to do with the slowdown: it points to pandemic-era over-hiring, rising interest rates that made new projects unprofitable, the mass migration of development to India, and, in the US case, tax changes eliminating deductions for programmer labor. With this cocktail, no company dares hire a junior; it’s not that they don’t produce work, but that they take months to become profitable.

The opposing thesis doesn't deny the above but adds a new layer. AI raises the floor: it gives juniors a notable productivity boost. The problem is it also removes their role as apprentices. Previously, newcomers acted as secretaries to seniors, handling repetitive tasks and soaking up the style and inner workings of the system. Today, a machine does that. The result is a generation swallowing generated code without knowing exactly what it does.

The ceiling remains human: what the MIT report says

A study led by MIT’s CSAIL team puts a brake on the euphoria. AI can provide code, solve LeetCode-style exercises, or alleviate heavy tasks, but large-scale programming is another world: maintaining systems with millions of lines, migrating obsolete technology between languages, testing faults only an expert eye recognizes, and documenting decisions in projects evolving for years.

Armando Solar-Lezama, one of the professors behind the report, summarizes it thus: “Everyone talks about automation, about programmers being unnecessary… but the real ‘work’ is much more complex than putting lines of code on the table.” The practical consequence is that the quality ceiling hasn't moved: good code still comes from a senior, or a senior supervising AI with clear instructions. For routine tasks, like writing tests, the tool works perfectly. For everything else, it doesn't.

Vibe coding and the coming technical debt

Meanwhile, a practice dubbed vibe coding has emerged: letting AI generate entire apps based on prompts. In discussion groups, some claim complete simple programs are already being built that work, with backend integration remaining the pending challenge. Enthusiasm has its downside. The lingering question is what happens when that software turns four, performance degrades, and someone opens the hood: if spaghetti code is inside, someone else will pay the bill.

According to this analysis, the market response has three predictable outcomes. Freezing the code and dragging legacy with patches until it becomes unviable. Hiring expensive teams to refactor what the machine generated. Or declaring the application obsolete and starting from scratch, likely again with AI, entering a cycle of new apps every two years. The full calculation, including cost scenarios, is where things get interesting.

From security engineer to bartender: the unexpected case

The hit isn't limited to juniors. There are two cases of security engineers who left the sector: one runs a bar, another works as a waiter. Both agree on the diagnosis: AI now performs pentesting, script creation, and documentation generation better than they do, salaries have dropped, and with tips, they earn the same without the burden of being available 24/7. It’s anecdotal, not statistical, but it points to displacement beyond entry-level roles.

The macro scenario suggests up to 90 million jobs lost to automation by 2030, against more than 170 million new ones. The net balance is positive in aggregate. The problem is that those losing their jobs don’t eat from the surplus. And some directly dismiss the optimistic part of the forecast: destroyed jobs are counted one by one, created ones are just promises.

What skills will be paid for in 2025?

Consensus, even among critics, is that the programmer profile is changing. They are no longer measured by how much code they write, but by how they debug and optimize it, prevent app obsolescence, and bridge business needs with technical requirements. Controlling AI, understanding where it fails, and not blindly depending on it becomes the main task. Those delegating without criteria generate problematic code.

The issue is universities are lagging. AI content arrives late in curricula, leaving updates to the student. Meanwhile, the market rewards those who already know how to integrate these tools and punishes those with only a degree. The question isn't whether AI will replace programmers. It's how many current CS students will actually practice before the sector finishes reordering itself.