AI news
AI model reviews
AI tools
How-to guides
Our staffAboutContact

Will AI replace software engineers? What the evidence says

AI now writes most new code at some big firms, yet developer jobs are still projected to grow. The real squeeze is on entry-level hiring.

Person with headphones working at a laptop
Photo: Julio Lopez on Unsplash

In April, Sundar Pichai told investors that about 75% of Google's new code is now generated by AI and then approved by engineers. That sounds like the end of the job. It isn't, at least not yet. AI isn't replacing software engineers as a profession, but it's already shrinking the number of junior roles, and the work itself is shifting from typing code to directing and checking it.

So the answer to "will AI replace software engineers?" splits in two. For experienced engineers, the data points to a changed job rather than a lost one. For people trying to get their first job, I think the worry is justified.

Will AI replace software engineers: in brief

No, not wholesale, and probably not soon. AI coding tools now write a large share of new code at big tech firms, yet US software developer employment is still projected to grow. The damage so far is concentrated at the entry level, where hiring of 22 to 25 year olds in AI-exposed jobs has fallen well behind everyone else.

AI coding tools have moved from autocomplete to agents

A few years ago, "AI for coding" meant a tool that suggested the next line while you typed. You accepted it or you didn't. The human still decided what to build, how to structure it and when it was done.

The current tools work at a different level. Agents such as Claude Code, OpenAI's Codex and Google's Antigravity can take a task described in everyday language, read a codebase, edit several files, run the tests and try again when something breaks. Anthropic studied 500,000 coding conversations and found that 79% of Claude Code sessions were automation, meaning the AI did the task directly, compared with 49% on the regular Claude.ai chat. The more agentic the tool, the less the human types.

That's the background you need for the replacement question. Writing code was never the whole job (I'd guess it's maybe a third of most engineers' week), but it was the part that was easiest to measure. It's also the part that's now easiest to hand off. If you want the general version of this argument across other careers, our piece on AI agents at work covers it.

A historical comparison helps here, and I'll use just this one. When VisiCalc arrived in 1979, spreadsheets did in seconds the arithmetic that clerks had done by hand for days. Accounting didn't disappear; the hand-arithmetic part of it did, and the people who stayed spent their time on judgement, rules and clients.

Coding agents look a lot like that to me. They take the typing and leave the deciding.

Real examples from Google, Anthropic and the job data

Google is the clearest company example. The share of AI-generated new code there went from about a quarter in October 2024 to roughly half by the end of 2025, then to 75% by April. In his first-quarter remarks, Pichai described the shift this way: "Our engineers are now orchestrating fully autonomous digital task forces, and building at a faster velocity." Google still employs engineers to approve all of that code, which is worth keeping in mind whenever you see the 75% figure quoted on its own.

The labour market data is more sobering. The Stanford Digital Economy Lab's "Canaries in the coal mine" paper, revised in August using ADP payroll data through June, found no evidence of widespread job displacement across the economy. But employment of workers aged 22 to 25 in AI-exposed occupations, a group that includes software developers, now sits 19% below where it would be if it had kept pace with less-exposed peers. Experienced workers show no comparable gap.

Two details in that paper matter more than the headline. The gap comes mainly from companies hiring fewer young people, not from firing them. And it's concentrated in jobs where AI substitutes for tasks; where AI complements workers, employment is flat or rising. The authors call these early, descriptive signals rather than proof that AI caused the drop, and I think that caution is fair.

Then there's the long-term outlook. Government projections listed on O*NET put US software developer employment at 1,717,800 in 2025, rising to 1,892,600 by 2035. That's 10% growth, with about 95,300 openings a year. Projections can be wrong, of course, and these were made with AI already in the picture.

SignalWhat it showsPoints towards
Google's new codeAbout 75% AI-generated, reviewed by engineersA changed job
Claude Code usage79% of sessions are direct automationFewer hands-on coding tasks
Stanford payroll studyYoung workers in exposed jobs 19% behind; seniors notFewer entry-level roles
US projections to 203510% growth, about 95,300 openings a yearNo collapse in total jobs

Benefits and limitations of AI coding agents

The benefit everyone talks about is speed. Pichai said a complex code migration done by agents and engineers together finished six times faster than engineers alone could manage a year earlier. Migrations, test writing and boilerplate are exactly the dull, well-specified work agents handle best.

The limitation is that speed is hard to prove outside flagship projects. METR ran a randomised trial at the start of 2025 and found experienced open-source developers took 19% longer on tasks when they were allowed to use AI. When it tried to repeat the study, the follow-up fell apart for an odd reason: many developers wouldn't agree to work without AI, and 30% to 50% admitted they'd held back tasks they didn't want to do by hand. METR now thinks developers are probably faster than in 2025, but it can't say by how much.

That's a strange position for an industry to be in. People are certain the tools help, and the best controlled evidence we have can't confirm the size of the gain.

There are other limits too. Anthropic's data shows the most common work handed to Claude was web front-end code in JavaScript and HTML, and its researchers suggested that jobs centred on "making simple applications and user interfaces" may feel the pressure sooner than backend roles. Agents still struggle with vague requirements, messy legacy systems and decisions about trade-offs, which is where senior engineers spend most of their day. Someone also has to own the code when it fails at 3am (an agent can't be on call in any meaningful sense).

Practical takeaways if you write software

If you're already mid-career, I'd stop asking whether you'll be replaced and start asking which of your tasks an agent could do next quarter. Then move your time towards the parts it can't: system design, reviewing AI output carefully, talking to the people who need the software. Getting good at specifying work clearly is now a coding skill in its own right.

Juniors have the harder road. The Stanford numbers suggest the first rung is where companies are cutting, so a CV that shows you can ship working products with agents, and explain and fix what they produce, probably counts for more than one listing languages. Free tools make that cheap to practise; our list of free AI coding assistants is a reasonable starting point, and the Cursor review covers one of the paid editors many teams use.

Students choosing a degree should look at the whole picture. Software is still projected to grow, and our guide to jobs that are safe from AI shows how few careers are untouched anyway. Honestly, I'd still pick computer science over most alternatives, but I'd pair it with a field you can build software for, such as health, finance or engineering.

The next few years for software jobs

What we know: AI writes a growing share of code at large firms, junior hiring in exposed jobs has dropped, and total developer employment hasn't fallen off a cliff. What we're guessing: whether the junior squeeze is temporary or permanent.

My own guess is that it's partly cyclical and partly structural. Companies that stop hiring juniors entirely will, in five or six years, run short of mid-level engineers who know their systems; that tends to correct itself. But I don't expect entry-level hiring to return to the 2021 pattern. The job of a junior engineer will likely look more like a reviewer and tester of agent output than a writer of tickets from scratch.

I'd watch two things. One is the Stanford team's tracker, since the gap for young workers has widened every time they've updated it. The other is whether studies like METR's can finally measure real productivity gains, because if they turn out large, the pressure on headcount will grow. For a broader view across professions, see our explainer on AI taking over jobs.

FAQ

Will AI replace engineers?

Not in general. AI is good at well-defined tasks like writing code to a clear spec, but engineering also means judgement, trade-offs, safety sign-off and responsibility for failures. Those parts are hard to automate. The bigger risk is fewer entry-level openings, since firms are hiring fewer juniors in AI-exposed roles while senior employment holds steady.

Will software engineers be replaced by AI?

Probably not as a profession. US projections still show developer jobs growing 10% between 2025 and 2035. What's changing is the work itself: at Google, about 75% of new code is AI-generated and reviewed by engineers.

Will AI replace engineering jobs?

It'll replace some tasks within engineering jobs rather than whole jobs, at least for now. Stanford's payroll study found no widespread displacement across the economy, but employment of 22 to 25 year olds in AI-exposed occupations is 19% behind where it would otherwise be. Young workers in exposed fields carry most of the risk so far, while experienced engineers show no comparable gap.

Will AI replace software engineers in 2026?

No. Companies are using agents heavily this year, yet they still need people to plan, review and ship the code. The real effect has been slower hiring of graduates, not mass layoffs. If you're job hunting, expect tougher competition for junior roles.

Can AI replace software engineers?

AI can already write a lot of software on its own, especially simple web apps and user interfaces. It still can't reliably turn vague business needs into working systems, weigh trade-offs or take responsibility when production breaks. For small, well-specified projects it gets close; for real products built by teams, it doesn't, as far as I can tell.

Keep reading