Will AI replace programmers?

The question is older than you might think. When GitHub Copilot first appeared in 2021, people asked whether autocomplete for code would make programmers unnecessary. Today, coding agents can take a task, edit dozens of files, run the tests and open a pull request on their own. The question is more serious now, and the honest answer is more interesting than yes or no.

What the evidence shows so far is a change in what programmers do rather than a disappearance of the role. Agents handle more of the typing. People still decide what should exist, notice when something is subtly wrong, and answer for the result when it breaks.

What does the research say about productivity?

The most careful study so far comes from METR, a research organization that measures what AI systems can do. In a randomized trial run in early 2025, 16 experienced open-source developers worked on 246 real issues in their own projects, with each issue randomly assigned to "AI allowed" or "AI not allowed." With AI, mostly Cursor with Claude 3.5 and 3.7 Sonnet, they took 19% longer. Yet they had expected AI to make them 24% faster, and even afterward believed it had sped them up by 20%.

In February 2026, METR published early results from a follow-up with 57 developers and more than 800 tasks, using newer agent tools. The picture flipped: returning developers were estimated to be about 18% faster with AI, though the uncertainty is wide. METR itself calls this weak evidence. A growing number of developers refused to take part because they did not want to work without AI, and many held back tasks they didn't want to do by hand, which probably makes the measured speedup too small. METR also mentions that about 4% of GitHub commits are now authored by Claude Code alone.

Two lessons follow. First, AI's effect on real programming work changed quickly between 2025 and 2026, and the tools are now part of normal practice for many developers. Second, how fast you feel is a poor guide to how productive you are, which is why measuring and reviewing still matter.

What about jobs?

The US Bureau of Labor Statistics (BLS) looked at this directly in projections published in early 2025, based on information available in mid-2024. It calls AI's effect on software developers highly uncertain, yet still projects their employment to grow 17.9% from 2023 to 2033, against 4.0% for the economy as a whole. Its reasoning: cheaper software tends to increase demand for software, and AI systems themselves need people to build and maintain them. BLS expects the sharpest declines in jobs whose core tasks a language model can copy easily, such as medical transcription and some customer service work, and says it will revise as AI evolves.

Projections are not promises, and averages hide differences between roles and regions. But neither the job data nor the productivity studies point to programmers simply being replaced. They point to fewer hours spent writing routine code and more weight on judgment.

A worked example: the same task, two ways

Here is how one small feature, "let customers reset their password," looks in 2026.

The agent did most of the typing. Every row that decides whether the feature is safe still depends on a person who understands software.

Which skills should learners build now?

Focus on the parts of the work that agents depend on people for:

• Reading code. You will review far more code than you write. Practice explaining, line by line, what a function does, and ask the AI to quiz you on its own output.
• Specifying clearly. Writing requirements, acceptance criteria and small task lists is now a daily skill, because an agent builds what you describe, not what you meant.
• Testing and debugging. Tests are how you prove an agent's work is right. Learn to write a test that fails for the right reason, and to find the cause of a bug instead of asking for random fixes.
• Security and data basics. Logins, permissions, secrets and personal data are where AI-written code fails most often and where mistakes cost the most.
• Git and tooling. Branches, pull requests and reverting a bad change let you work safely alongside agents.
• Understanding the problem. Knowing a field, whether accounting, health or logistics, is what lets you spot that correct-looking code does the wrong thing.

A useful way to practice is to ask the AI to teach rather than do:

I am learning to program in Python. Give me a small task, but do not write the solution.
When I send my code, review it like a senior developer: point out bugs, security issues and missing tests, ask me questions that make me explain my choices, and only show a corrected version after I have tried to fix it myself.

Set the review bar high from the start: ask for explanations, run the tests yourself, and keep a short list of the mistakes you catch. That list shows you which skills to work on next.