Will AI replace programmers? What the benchmarks actually say
Status as of 2026-09-05
- BLS Employment Projections 2024–34: computer programmers −6%, software developers +15%, data scientists +33.5% — the same field, opposite directions by occupation.
- Indeed Hiring Lab (2026-07-08): 71% of software-development posting growth May 2025→May 2026 was senior-level; the most AI-exposed occupations’ postings fell the most and then rebounded the most.
- As reported by secondary coverage of the NY Fed’s “Labor Market for Recent College Graduates”: unemployment 6.1% for computer-science majors and 7.5% for computer engineering — a junior-hiring problem, not a profession-wide one. Full reconciliation in the five-source card below.
Start here if the question is about your own job: Should you be worried about AI taking your job? →
What AI can already do
The measurable progress is real. Frontier models handle complex, multi-step coding tasks — around 80% on SWE-Bench Pro — and coding agents are deployed in real production workflows, not just demos. On raw capability, AI has cleared the bar of a strong entry-level engineer for many well-scoped tasks. This is the same evidence behind the scorecard's On-track "outpace college graduates" verdict.
Where it still falls short
Benchmarks aren't autonomy. The gap between scoring 80% on a coding test and reliably owning a feature end-to-end — scoping, integrating, debugging in a messy real codebase, and being accountable for it without human cleanup — is exactly where "replace" breaks down. The defining AGI bar Aschenbrenner named, a system that autonomously does AI research (the hardest kind of programming), remains undemonstrated.
The honest verdict
| Question | Answer (mid-2026) |
|---|---|
| Can AI write production code? | Yes, with review |
| Does it automate parts of the job? | Yes |
| Can it replace an engineer unsupervised? | Not yet |
So: AI is automating slices of programming and raising each engineer's output, but it is not a drop-in replacement for a software engineer as of mid-2026. On track on capability; the reliability-and-ownership gap is what to watch.
Five sources, one card
The programmer debate runs on five numbers that sound contradictory. Here they are side by side, each dated and linked.
| Source | What it measures | Reading (dated) | Link |
|---|---|---|---|
| BLS Employment Projections 2024–34 | Projected US employment change by occupation, ten years out | Computer programmers −6%; software developers +15% (data scientists +33.5%). Published 2026. | bls.gov |
| NY Fed, “Labor Market for Recent College Graduates” | Unemployment of recent graduates by major | As reported by secondary coverage: 6.1% computer science, 7.5% computer engineering (2026). | Secondary reports; not independently verified here |
| Indeed Hiring Lab | Job-posting growth by seniority | 71% of software-development posting growth May 2025→May 2026 was senior-level (2026-07-08). | hiringlab.indeed.com |
| Microsoft Research, “Working with AI” | AI applicability: share of an occupation’s work activities Copilot users were observed doing with AI | Computer Programmers (SOC 15-1251) 0.31; Software Developers (SOC 15-1252) 0.278 (Jul 2025, CC-BY-4.0). | github.com/microsoft |
| Anthropic Economic Index | Share of occupations’ tasks observed in Claude usage; augmentation vs automation | 49% of occupations have ≥25% of tasks observed; usage split 55% augmentation vs 42% automation (March 2026). | anthropic.com |
Why they disagree: exposure ≠ displacement. Microsoft and Anthropic measure how much of the work AI is observed doing; the BLS projects headcount a decade out; Indeed and the NY Fed measure hiring and unemployment now. An occupation can be highly AI-applicable and still grow (developers, +15%) while a neighbouring title shrinks (programmers, −6%), and a junior-hiring slump can coexist with senior growth (71%). Look up your own SOC code on the AI job exposure check.
Five sources, one card — so when does an agent own the whole engineering job?
One tap — see which real forecaster you side with, instantly. No sign-up.
Frequently asked questions
Partly, and unevenly. As of mid-2026 agentic coding is strong (~80% on SWE-Bench Pro) and automates parts of the job, but no system reliably owns end-to-end software engineering without human review. It augments programmers rather than replacing them.
Frontier models score around 80% on SWE-Bench Pro and ship real code in production for well-scoped tasks — roughly the level Aschenbrenner predicted for 2025/26, graded On track.
The gap between benchmark scores and reliable, accountable, end-to-end ownership in a real codebase. Autonomous software engineering without human cleanup is undemonstrated as of mid-2026.
The live scorecard updates as models ship and verdicts change.
View the live scorecard →Be first to know when a verdict flips
One email when one of the eight verdicts changes and the AGI-2027 Thesis Tracker moves — the single auditable score no other tracker has. Not a weekly newsletter; nothing arrives until the score actually moves.
Subscribe free →