Should You Still Learn to Code Now That AI Can Write It?

Every few weeks someone posts the same argument: AI writes the code now, so spending years learning to program yourself is a waste of time.

Nvidia's Jensen Huang has been making a version of this case since 2024, when he told the World Government Summit in Dubai that computing's whole job is to make programming unnecessary, since human language is now the only programming language anyone needs, according to TechRadar. I think he is wrong, and the evidence for that is not a hunch.

Google gives Huang's argument some real numbers to stand on. More than a quarter of its new code was AI generated by late 2024, CEO Sundar Pichai told investors on a call covered by Fortune, and that share has climbed since.

So the code volume argument is not made up. What it leaves out is who is checking that code, and whether that person still knows what they are looking at.

What happens when you actually measure it

That question got a real answer in January 2026, when Anthropic ran a randomized trial on 52 mostly junior developers learning a Python library none of them had touched before. Half got AI help, half did not. Both groups then took a quiz on debugging and comprehension with no AI allowed.

The AI group finished the coding task barely faster and scored 50 percent on the quiz. The unaided group scored 67 percent, with the widest gap on debugging, per Anthropic's own writeup and the arXiv paper behind it.

Something similar showed up in METR's study of experienced open source developers a few months earlier. Given real issues on repositories they had worked in for years, developers using AI tools took 19 percent longer than developers working without them.

Before starting, they had guessed AI would make them 24 percent faster. Afterward, they still believed it had, estimating a 20 percent speedup that never happened. Feeling faster and being faster turned out to be two different things, off by close to 40 points.

GitClear's research on actual codebases backs this up from a different angle. Across more than 200 million changed lines, refactoring dropped from about a quarter of all changes to under 10 percent, while duplicated code blocks jumped eightfold, right as AI assisted commits became routine.

GitClear's own CEO has pointed to a simple mechanism: AI tends to write a new function instead of reusing one that already exists, because it does not carry the context of the codebase around the way a developer does.

The market already believes this

Undergraduate computer science enrollment fell 3.6 percent in fall 2025 and another 8.4 percent the following spring, with graduate programs down 14 percent, according to National Student Clearinghouse numbers reported by Fortune.

Read that next to the Stanford Digital Economy Lab finding that employment for software developers aged 22 to 25 dropped 13 percent relative to older developers in the same jobs since late 2022, nearly 20 percent from its own peak. Fewer people are studying it, and fewer of the ones who do are getting hired into it.

You could read that as proof AI already won and there is no point learning a shrinking trade. I read it the other way.

The entry level job used to be where fundamentals got built on the clock, with a senior engineer catching your mistakes as you went. That job is disappearing faster than the fundamentals it used to teach. If nobody is going to hand you that training anymore, you have to go build it yourself, on purpose, before you need it.

Stack Overflow's 2025 survey of about 49,000 developers shows the tension in real numbers. 84 percent are using or planning to use AI tools, up from 76 percent a year before, and trust in the results has dropped to about a third of respondents.

Two thirds say the output is almost right but not quite. Nearly half say debugging it takes longer than writing the code themselves would have, per ADTmag's coverage of the same numbers. Developers keep reaching for the tool and keep not believing it.

Here is the part of the Anthropic study that actually matters more than the headline number: it was not AI use itself that predicted the score, it was how people used it.

Developers who asked the assistant to explain its reasoning, or questioned why a piece of code worked, scored close to the unaided group. Developers who just accepted what came out scored the worst. The tool was not the variable. Getting stuck first was.

flowchart TD
    A[Developer works with an AI assistant] --> B{How is it used}
    B -->|Asks why, checks the reasoning, gets stuck first| C[Judgment and debugging skill grow]
    B -->|Accepts the output and moves on| D[Skill formation stalls]
    C --> E[Can catch AI mistakes later, under pressure]
    D --> F[Cannot tell good output from bad]

So learn to code. Not because the market owes you a job for it, and not because Huang is lying about where AI is headed.

Learn it because somebody has to be able to tell when the code is wrong, and that skill is built by writing bad code yourself first, not by reading good code somebody else generated.

Build the thing without help once. Then use AI to go faster on the parts you already understand.