Will AI replace software developers? The evidence
AI now writes a large share of new code at some big companies, and coding is the single biggest use of AI chat tools. Measured productivity gains for experienced developers are smaller and less certain than the claims, and the clearest job effect so far is fewer openings for young developers rather than mass layoffs of experienced ones.
The job today
Task by task: what AI can actually do
In real use means deployed at scale; shown in tests or pilots means demonstrated but not yet routine; not yet means we found no credible evidence it does this reliably.
| Task | Status | Evidence |
|---|---|---|
| Writing new code | In real use | Google CEO Sundar Pichai said in Apr 2026 that AI generates three-quarters of the new code Google writes, up from 25% in 2024, with engineers supervising AI agents. (Semafor, Apr 2026) |
| Delegating multi-step coding to agents | In real use | Computer and math tasks were 35% of Claude.ai conversations in Feb 2026, and coding work is shifting to agentic API use, which Anthropic says 'may signal more imminent transformation' of these jobs. (Anthropic Economic Index, Mar 2026) |
| Fixing issues in large, familiar codebases | Shown in tests or pilots | METR's randomized trial found experienced developers took 19% longer with early-2025 AI tools; its Feb 2026 follow-up estimated a 4% speedup for new recruits (range -15% to +9%) and called the data unreliable. (METR, Feb 2026) |
| Large code migrations | Shown in tests or pilots | Pichai said one complex code migration took about one-sixth of the time it would have a year earlier (company claim, Apr 2026). (Semafor, Apr 2026) |
| Deciding what to build and owning ambiguous requirements | Not yet | Benchmarks like GDPval give models clear prompts and reference files; OpenAI notes real tasks are often ambiguous and its test does not capture building context or iteration. (OpenAI GDPval, Sep 2025) |
Jobs evidence so far
Labelled by who says AI was the reason: the company itself, reporters, or a study.
- Block cuts about 4,000 of 10,000+ staff (40%); Jack Dorsey wrote 'Intelligence tools have changed what it means to build and run a company' (company says AI; roles cut not specified) (SF Standard, Feb 2026)
- Stanford/ADP: employment of 22-25-year-olds in AI-exposed occupations is 19% below trend, mainly through reduced hiring; the pattern holds excluding tech firms (study) (Stanford Digital Economy Lab, Aug 2026)
- Indeed: US software development postings up almost 15% since Claude Code launched but still about 27.5% below pre-pandemic levels (study) (OfficeChai on Indeed Hiring Lab, Jul 2026)
- Microsoft offers voluntary buyouts to about 60,000 employees, framed by press as an AI-era overhaul; memo excerpts do not mention AI (reported as AI) (TechCrunch, Oct 2026)
What to do
- Get good at directing and reviewing agentsWriting specs, breaking work into tasks and reviewing AI-written code is becoming the core skill at companies where most new code is AI-generated.
- Own the parts benchmarks skipRequirements, architecture, security, production incidents and trade-off decisions are where ambiguity lives and tools are weakest.
- Measure your own speedupMETR found developers believed AI sped them up even when it slowed them down; time real tasks with and without tools.
- Watch junior hiring, not headlinesThe clearest signal so far is fewer entry-level hires; early-career developers should build visible, shipped work.
Questions people ask
Will AI replace software developers?
Not on current evidence. BLS projects 10% growth to 2035, and Indeed postings rose in 2025-26 though they remain well below pre-pandemic levels. AI is changing the work rather than removing the role.
Is it true AI writes most code at Google?
Google's CEO said so in Apr 2026 (75% of new code). That is a company claim; how much human review that code gets was not detailed.
Does AI actually make developers faster?
Unclear. A 2025 randomized trial found experienced developers 19% slower; a 2026 follow-up suggested modest speedups but METR called its own data unreliable.
Are junior developers hit hardest?
Stanford's payroll study finds young workers in AI-exposed jobs 19% below trend, mainly through fewer hires. The authors call this descriptive, not proof that AI caused it.
Which developer tasks are safest?
Defining requirements, system design, security and incident response, and accountability for what ships.
How we grade
Each task is graded 'deployed' only when a named organisation reports AI doing it in real work at scale, 'demonstrated' when it works in tests, benchmarks or pilots, and 'not yet' when we found no credible evidence AI does it reliably. For job losses we separate what a company said in its own words ('company says AI') from what reporters or analysts attributed to AI ('reported as AI') and from research ('study'); employment and outlook figures come from the BLS 2025-35 projections.
Sources
- BLS Occupational Outlook Handbook: Software developers, QA analysts and testers
- METR: early-2025 developer RCT (Jul 2025)
- METR: uplift update (Feb 2026)
- Semafor: Google says 75% of new code AI-generated
- Anthropic Economic Index, Mar 2026
- Stanford: Canaries in the Coal Mine
- OfficeChai: Indeed software postings
- SF Standard: Block layoffs
- TechCrunch: Microsoft buyouts
Spot an error or something out of date? Tell us; corrections are made and logged. Written with AI assistance from public sources; every fact checked and approved by Colin Balfe. General information, not legal, medical or financial advice.