Work and careers · The ruling
Can AI replace software engineers?
Verdict
The verdict
No, not on current evidence. AI now does a large share of everyday coding tasks, while people still decide what to build, check the output and answer for what ships.
No means: no, or not in a way you should rely on. How we rule

Short answer: AI tools now write, test and explain a lot of code, and most developers use them every day. The published evidence shows AI taking over tasks inside the job. It does not show the job itself disappearing, and none of it can tell you what happens next.
Tasks versus jobs: what AI coding tools do well today
A software engineering job is a bundle of tasks. You work out what is actually wanted, break it down, write code, read other people’s code, test it, fix it, deploy it and explain it to people who don’t care about the details. “Can AI replace engineers?” mixes two questions. Can AI do some of those tasks? Plainly, yes. Can it do the whole bundle without supervision and answer for the result? Nothing in the evidence below shows that.
The 2026 Stack Overflow Developer Survey shows where the tools have landed. Coding assistants and agents are now the most common use of AI among respondents (66%), and 73% use them daily. People hand over the work they can check. The top current use was generating code in a familiar area (69.3%), followed by debugging, troubleshooting or refactoring, answering straightforward technical questions and writing tests. Communication and design lag well behind. Only 35% and 49% of respondents found AI useful there, against more than 70% for implementation and debugging.
Trust is conditional. In the same survey, 48% said they trust AI when they can easily validate its answer, and only 16% trust it for most tasks that are not important work decisions. Stack Overflow’s own conclusion is that a human in the loop is the dominant way of working.
What published studies and surveys actually measure
Three kinds of evidence get quoted in this argument, and each answers a different question.
Surveys measure adoption and attitude. They show that developers use AI heavily. They don’t measure output, and they don’t measure jobs.
Controlled experiments measure speed on a defined task. In the GitHub Copilot experiment (Peng et al., 2023), recruited developers were asked to build an HTTP server in JavaScript as quickly as possible. The group with Copilot finished 55.8% faster than the group without it. That was one self-contained task.
METR’s randomised trial pointed the other way in a harder setting. In early 2025 it recruited 16 experienced developers who had worked for years on large open-source repositories, and gave them 246 real issues from those projects. When AI was allowed, they took 19% longer. They had expected a 24% speed-up, and afterwards they still believed AI had made them 20% faster. METR now marks those results as out of date. Its February 2026 update says developers are now likely sped up more than in early 2025, but that its newer data is unreliable. More developers declined to take part because they did not want to work without AI, and 30% to 50% of those surveyed said they held back some tasks for the same reason. METR is redesigning the study. The lesson to take from it is narrow: results depend heavily on the task, the codebase and the tools, and developers’ own sense of how much faster they are is a poor guide.
Labour-market analysis measures exposure. The Department for Education6’s 2023 report on the impact of AI on UK jobs ranks occupations by how much the abilities they need overlap with what AI can do. The report says it does not distinguish whether a job will be aided or replaced. Programmers and software development professionals do not appear in either of its top 20 lists of most exposed occupations. Those lists are led by roles such as management consultants, accountants, solicitors and psychologists. The report dates from November 2023, before today’s coding agents, so treat it as dated.
For comparison, here is some US data, which does not carry over to the UK. The US Bureau of Labor Statistics7 projects that employment of software developers, quality assurance analysts and testers will grow 10% from 2025 to 2035, against 3% for all occupations. It puts part of that demand down to software development for AI. That figure is a projection.
None of these sources measures whether a particular engineering job will still exist in ten years.
Where AI still needs a human engineer
Deciding what to build and running it in production. In the 2025 Stack Overflow survey, the tasks developers most often said they did not plan to use AI for were deployment and monitoring (76%) and project planning (69%).
Security-critical code. The NCSC says AI models are trained on code that includes security flaws. If you let an AI loose on your codebase with little oversight, there is a real risk it produces vulnerable code. The NCSC’s advice is to match oversight to risk. A prototype or demo can be mostly AI-written. Authentication, sensitive personal data, secrets and credentials, and anything safety-critical need you to review the code, understand it, check it for vulnerabilities and verify that it does what you expect. It expects trust in AI-written code may grow as models improve, but says “we’re not there yet”.
Owning the result. When code breaks in production, a person or an organisation has to answer for it. The tools change who types the code. They don’t change who is responsible for it.
What this means for UK entry-level and junior roles
Editorial note: the sources above do not settle long-term demand for engineers, what future models will be able to do, or the effect on junior hiring. We found no UK primary data on how AI has changed junior software hiring, so this section gives no figures.
The tasks AI handles best overlap with the smaller, well-defined work juniors have traditionally learned on, so the worry is reasonable. The evidence also points both ways. The Copilot paper’s authors saw promise in AI helping people move into software careers. In the 2026 Stack Overflow survey, the most common reason respondents gave for avoiding AI at work or school was concern over losing job skills or training an AI to replace them (17.1%).
- Students: check your course’s rules before using AI on assessed work. Use it to understand code, not to hand in code you can’t explain (help me study covers how).
- Job hunters: show work you can explain line by line. See how to write your CV and answer interview questions.
- Employers: give staff approved AI tools, and set review rules by risk, as the NCSC’s spectrum suggests.
How to do it well
- Start with work you can check. Use AI for tests, refactors, explanations and code in areas you already know. That is where the survey shows developers are most confident.
- Read every line before you merge. If you can’t explain why the code works, you’re not done.
- Write the tests yourself for anything important, then let the AI write code until they pass. The NCSC names this test-driven approach as one middle ground between manual coding and full “vibe coding”.
- Set oversight by risk. Give it a free hand on prototypes and internal demos. Keep tight control over authentication, personal data, credentials and safety-critical code.
- Keep your own skills sharp. Do some work without help, and ask the tool to explain its code as well as write it.
If you aren’t a developer and want to build something, create an app and build a website show where AI helps and where you still need to understand what it made.
Rules and risks
Confidential code and data. The NCSC warns that information you put into consumer AI services may be stored, retained or used to improve the service unless specific privacy controls are in place. Your employer can also lose visibility and control of it. Use the tools your organisation has approved, and don’t paste in secrets, credentials or customer data.
Who owns AI-generated code. Under section 9(3) of the Copyright, Designs and Patents Act 1988, the author of a computer-generated work is the person who made the arrangements necessary for its creation. Section 178 defines “computer-generated” as made by computer where there is no human author. The Act doesn’t say how much human input makes you the author of code you wrote with AI help. If ownership matters commercially, take legal advice.
Security. Treat generated code as unreviewed until a person has reviewed it, especially anything on the NCSC’s high-risk list.
This page is general information, not legal or careers advice.
Sources
11- Stack Overflow Developer Survey 2026: AI
- Stack Overflow Developer Survey 2025: AI
- METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
- METR: We are Changing our Developer Productivity Experiment Design (February 2026)
- Peng et al. (2023): The Impact of AI on Developer Productivity: Evidence from GitHub Copilot
- Department for Education: The impact of AI on UK jobs and training
- US Bureau of Labor Statistics: Occupational Outlook Handbook, Software Developers, Quality Assurance Analysts, and Testers
- NCSC: The 'vibe coding spectrum' approach to AI-assisted software development
- NCSC: The hidden risks of shadow AI
- Copyright, Designs and Patents Act 1988, section 9
- Copyright, Designs and Patents Act 1988, section 178
FAQ
Questions people ask
Will AI replace software engineers in the next 5 to 10 years?
No published source can tell you. The studies and surveys measure how developers use AI today and how fast they work on particular tasks. None of them measures whether engineering jobs will exist in a decade. Long-term demand and what future models will be able to do are still open questions, so treat any confident forecast either way as a guess.
Is it still worth learning to code or studying computer science in the UK?
We can't give you a hiring forecast, because the sources don't settle what will happen to junior roles. What they do show is that people use AI most where they can check the answer. The 2026 Stack Overflow survey found 48% of respondents trust AI when they can easily validate its output. Knowing how software works is what lets you do that checking. Before you use AI on coursework, check your course's rules.
Which software engineering tasks can AI already do?
In the 2026 Stack Overflow survey, the most common current uses were generating code in a familiar area, then debugging, troubleshooting or refactoring, answering straightforward technical questions, writing or improving tests, and generating code in an unfamiliar area. Far fewer respondents found it useful for communication and design.
Is AI-generated code safe to ship without review?
Not for anything that matters. The NCSC says letting AI loose on a codebase with little oversight carries a real risk of security vulnerabilities. It advises more human control for authentication, sensitive personal data, secrets and credentials, and safety-critical systems. That means reviewing the code, understanding it, checking it for vulnerabilities and verifying that it does what you expect.
Do junior developers face more risk than senior ones?
Possibly, but no source we checked measures it. The tasks AI handles best overlap with work traditionally given to juniors, so the concern is reasonable. A reasonable concern is still not a documented result, and we found no UK primary data on how AI has affected junior software hiring.
General information for UK readers, not legal, financial or medical advice. How verdicts are decided.