The short answer. Not the job, but parts of it. BLS projects software developer employment to grow 10% from 2025 to 2035, while computer programmer jobs are projected to fall 7%. Controlled studies found measurable gains on everyday coding and, in early 2025, a slowdown for experienced developers in large codebases. Entry-level hiring is where the pressure shows.
AI assistants now complete lines, write tests and, increasingly, make whole changes as agents. So will AI replace software engineers? The best evidence says it reshapes the work unevenly: gains on routine code, mixed results in big codebases, and fewer openings for juniors.
A note on names: the US Bureau of Labor Statistics (BLS) tracks software developers (who design and build software) separately from computer programmers (who mainly write and test code). That split matters here.
What does AI already do in software engineering work?
AI writes a growing share of routine code: completions, boilerplate, tests and small changes. In large field trials it raised developers' completed tasks by about a quarter. In mature open-source codebases with high quality bars, early-2025 tools slowed experienced developers down. Design, review and ownership of what ships stay with engineers.
| What AI does today | What stays with the engineer |
|---|---|
| Code completion and boilerplate. Across three randomized trials with 4,867 developers, access to an AI coding assistant raised completed tasks by 26.08% (standard error 10.3%) | Deciding whether the code belongs in the system at all |
| Changes in large, mature codebases. In METR's 2025 trial, experienced open-source developers took 19% longer with early-2025 AI tools | The deep context and quality bar of a codebase you know well |
| Repetitive programming tasks. BLS expects companies to use AI "to automate repetitive programming tasks" | Higher-skilled work, which BLS says will likely "shift to other workers, such as software developers" |
| Agentic changes across a repository. Use of agentic coding tools rose among open-source developers through 2025, METR reports | Defining the task, reviewing the result and owning what merges |
| Requirements and system design. Drafts options, diagrams and documentation on request | BLS lists the first duty as "Analyze users' needs and then design and develop software to meet those needs" |
What the speedup studies measured
The largest controlled evidence comes from randomized trials at three large companies, run as part of normal business and now published in Management Science. A random subset of 4,867 developers got an assistant that suggested code completions, and completed tasks rose 26.08% across the pooled trials.
The limits matter. The authors write that "each experiment is noisy and results vary across experiments," the tool predates today's agents, and the outcome counts completed tasks rather than long-term code quality. They also found that "less experienced developers had higher adoption rates and greater productivity gains."
What the slowdown study measured
METR's randomized trial went the other way. Sixteen experienced developers worked on 246 real issues in large open-source projects they had contributed to for years. With AI allowed, tasks took 19% longer, yet afterwards the developers believed AI had sped them up by 20%.
That gap between feeling faster and being faster matters for any engineering team. The result has limits too: 16 developers, one kind of codebase, early-2025 tools. In a February 2026 update, METR said it now believes "it is likely that developers are more sped up from AI tools now" than in early 2025, but that its newer data "is only very weak evidence for the size of this increase," partly because many developers no longer wanted to work without AI.
Controlled studies of AI coding tools point in different directions, partly because they measure different work. Trials with 4,867 developers found 26% more completed tasks with a code-completion assistant. A 2025 trial of 16 experienced open-source developers found tasks took 19% longer, even though the developers believed they were 20% faster.
What does the job outlook say for software engineers?
The outlook splits by role. BLS projects software developer, QA analyst and tester employment to grow 10% from 2025 to 2035, against 3% for all occupations, with a median wage of $135,980 for software developers in May 2025. Computer programmer jobs are projected to decline 7%, and BLS names AI as one reason.
For developers, BLS writes: "Overall employment of software developers, quality assurance analysts, and testers is projected to grow 10 percent from 2025 to 2035, much faster than the average for all occupations." It also expects demand to be strong partly "due to the continued expansion of software development for artificial intelligence (AI), Internet of Things (IoT), robotics, and other automation applications."
For programmers, the tone is different: "Computer programming work continues to be automated, helping computer programmers to become more efficient in some of their tasks."
| Occupation | Change, 2025-35 | Openings a year |
|---|---|---|
| Software developers, QA analysts and testers | +10% | About 106,100 |
| Computer programmers | -7% | About 4,400 |
Programming is also a much smaller occupation: 110,800 jobs in 2025 against 1,905,400 in the developer group, with a May 2025 median wage of $100,390.
The pressure is sharpest at the start of careers. Stanford Digital Economy Lab researchers, using payroll data through June 2026, found "no evidence of widespread, economy-wide job displacement." But employment of workers aged 22 to 25 in AI-exposed occupations "now stands 19% below where it would be had it kept pace with that of their less-exposed peers," mainly through less hiring. Their tracking dashboard shows "large declines for early-career workers" among software developers. The authors describe these as early, descriptive indicators, not causal estimates.
Which skills keep software engineers valuable?
The skills that gain value sit around the code: understanding what users need, designing systems, reviewing what AI produces and directing agents with good context. Typing routine code matters less. Judging whether code is correct, safe and maintainable matters more.
- System design and requirements. BLS describes developers as people who "Design each piece of an application or system and plan how the pieces will work together." AI can propose a design. Choosing one that fits your constraints is still engineering.
- Review and verification. METR counted a task done only when the developer was satisfied the code would pass review, including style, testing and documentation. That bar is where human review earns its keep, and knowing how models produce plausible but wrong output makes the review faster.
- Context engineering for coding agents. An agent with the wrong context writes the wrong code quickly. Giving it constraints, conventions and examples is a skill, covered in context engineering vs. prompt engineering.
- Directing and supervising agents. As tools move from suggestions to multi-step changes, engineers spend more time scoping tasks and checking results. kju's AI agents track and our guide to agentic AI skills cover how.
- Measuring instead of guessing. METR's developers felt faster while being slower. Tracking cycle time and defect rates, rather than impressions, is the simplest guard against the AI productivity paradox.
Try three real kju questions for software engineers
Here are three from kju's daily practice for engineering teams at software companies, across difficulty levels. kju users answer eight like these every day, and the set adapts to what they miss.
Question 1. You're using AI to produce a code review summary. The first output is generic and misses the point. Best next move?
- A) Give it context: audience, what good looks like, an example
- B) Regenerate until it comes out right
- C) Give up: the model can't do this task
- D) Ask it to make the answer longer
Answer: A. Models aren't mind readers: context, a definition of good, and an example beat blind regeneration. Source
Question 2 (true or false). To save time, it's fine to paste customer usage data and source code into any free public AI chatbot.
- True
- False
Answer: False. Public chatbots may store or train on what you paste. Use your company's approved tools for anything sensitive. Source
Question 3. Which are real failure modes to design for in production LLM systems? Select all that apply.
- A) Injection via untrusted input
- B) Fabricated citations
- C) Silent truncation of long context
- D) Learning from each chat by default
Answer: A, B and C. Injection, fabricated citations and context truncation are real; models don't permanently learn from your chats by default. Source
What should engineers and engineering leads do next?
If you're an engineer, run your own small experiment: time a few comparable tasks with and without AI, and note what you had to fix. METR's results show why your impression alone isn't enough. Then build the review and context habits with a few minutes of daily practice on kju.
If you lead an engineering team, the open questions are about standards: which tasks go to agents, how AI-written code gets reviewed, and how juniors still learn the system. kju for teams gives every engineer daily practice tied to that work, and our page for technology and engineering teams shows what that covers.
For the wider view, read what jobs AI will replace, or see how the question plays out for data analysts and project managers.
Frequently Asked Questions
- Will AI replace programmers?
- Some programming work, yes. BLS projects employment of computer programmers to decline 7% from 2025 to 2035 and says companies are expected to use AI to automate repetitive programming tasks. Software developers, who design systems as well as write code, are projected to grow 10% over the same period.
- Will engineering be replaced by AI?
- There is no sign of that in the data so far. Stanford researchers using payroll data through June 2026 found no evidence of widespread, economy-wide job displacement. They did find that employment of 22- to 25-year-olds in AI-exposed occupations, including software developers, stands 19% below where it would be had it kept pace with less-exposed peers.
- Does AI make software developers more productive?
- It depends on the task and the setting. Three randomized trials with 4,867 developers found a 26.08% increase in completed tasks with an AI coding assistant. A smaller 2025 trial found experienced open-source developers took 19% longer with early-2025 tools, though its authors now believe developers are likely faster than that.
- Is it still worth learning to code in 2026?
- BLS still projects about 106,100 openings a year for software developers, quality assurance analysts and testers, and 10% growth from 2025 to 2035. What changes is the mix: less typing of routine code, more design, review and directing AI tools. Coding knowledge is what lets you judge whether AI-written code is right.