Will AI Replace Computer Programmers?
The single highest observed AI coverage of any occupation in Anthropic's data — AI writes code all day, every day. But note what BLS is shrinking: the narrow "programmer" occupation of writing code to specification, while software developers, who design systems, are projected to grow.
The exposure evidence — three sources, side by side
There is no official "AI risk score" for any job, and any site that gives you one number without a source is making it up. Here are the three serious research measures that exist for this occupation, each with what it actually counts:
| Research source | Exposure figure | What it actually measures |
|---|---|---|
| Anthropic Economic Index — observed exposure | 75% of tasks | Share of this occupation's tasks where AI use was actually observed in real Claude conversation data (Anthropic, "Labor market impacts of AI", March 2026 dataset). This is measured use, not a prediction. |
| GPTs are GPTs (OpenAI / University of Pennsylvania) — GPT-4 rating, β measure | 95% of tasks | Share of tasks rated exposed to large language models, counting direct exposure in full and assisted exposure at half weight (the paper's β measure). Human expert raters scored the same occupation at 68%. |
| Microsoft "Working with AI" — AI applicability score | 0.31 / 1.00 | A 0–1 score built from 200,000 real Bing Copilot conversations, weighting how often AI was used for this occupation's tasks and how well it completed them. It is a score, not a percentage; the median occupation scores about 0.14. |
The three sources measure different things — observed use, rated capability and Copilot-based applicability — so they rarely agree exactly. Where they disagree, that disagreement itself is information: it usually means a job is being changed by specialist tools, or in ways general chatbots do not show. All three agree on the framing: exposure is not replacement.
The work itself: what AI can help with, and what stays human
These are this occupation's real tasks, from the US Department of Labor's O*NET database (version 30.0). The split is our editorial analysis — the tasks on the left are the routine, screen-based ones current AI tools demonstrably assist with; the ones on the right involve physical presence, licensed responsibility, or judgement about people.
Where AI can already help
- Writing, updating and maintaining programs for specific jobs such as inventory or data retrieval
- Correcting errors by changing code and rechecking the program
- Compiling and writing documentation of program development and revisions
- Preparing workflow charts and diagrams of program logic
What stays human
- Clarifying intent with managers, engineers and users before code is written
- Systems analysis: deciding what should be built and how it fits together
- Investigating why a whole system — network, hardware, software — misbehaves
- Directing and reviewing the work of other programming personnel
Pay, outlook and entry requirements (BLS)
| Median annual pay, 2024 | $98,670 |
|---|---|
| Employment change, 2024–34 | BLS projects employment to shrink by 6% from 2024 to 2034 (121,200 → 113,900 jobs, a change of -7,200). |
| Typical entry-level education | Bachelor's degree |
Basis: BLS Occupational Outlook Handbook, 2024–25 edition profile (2024 pay, 2024–34 outlook). A median is the midpoint of what all workers in the occupation earn — it is not a starting wage, and it is not a forecast.
What to learn next
A roadmap only helps if it starts from the skills this job already builds. These three moves do — each is chosen because the destination work is growing, harder to automate, or carries a credential software cannot hold:
- AI-assisted development workflows → Software Developers. The growth is in the developer role — design, architecture and product judgement — which is projected to grow while code-to-spec work shrinks. See: Software Developers — projected +15% 2024–34 (BLS).
- Testing, QA and code review → a stronger position in the same field. More AI-generated code means more demand for people who can verify it: testing strategy, security review and debugging production systems.
- A specialism with a moat — security, data or embedded systems → a stronger position in the same field. Deep domain contexts (regulated industries, hardware, safety-critical code) are where generic code generation is weakest and experience is worth most.
AI tools in this job today
AI coding assistants now draft, complete and debug routine code, and agentic tools run trial-and-error loops on their own. The programmer tasks most exposed are the well-specified ones; the durable tasks involve understanding a business problem, a legacy system and the people who depend on both.
Frequently asked questions
Will AI replace computer programmers?
Not on any schedule the research supports. The strongest exposure figure for this occupation is very high — about 95% of its tasks are within AI's reach — but exposure measures tasks, not whole jobs. BLS projects this occupation shrinking 6% between 2024 and 2034, and Anthropic's own labour-market research found no systematic rise in unemployment among workers in highly exposed occupations. The realistic outcome is a changed job, not an abolished one.
Which parts of a job are most exposed?
The tasks listed under “Where AI can already help” above — mostly the routine, text- and data-handling parts of the role. They come from this occupation's real O*NET task list, and they are the parts to stop competing on and start supervising.
How much do computer programmers earn?
The BLS median annual pay was $98,670 in 2024 — half earned more, half earned less. It is a national median, not a starting salary or a promise; pay varies widely by state, employer and experience.
What should I learn next?
The roadmap above is built from this occupation's actual skills: AI-assisted development workflows; Testing, QA and code review; A specialism with a moat — security, data or embedded systems. Each points at work that is either growing, harder to automate, or both — with the BLS outlook for the target occupation stated next to it.
Where do these numbers come from?
Exposure figures come from three published research datasets — Anthropic's Economic Index, the OpenAI/University of Pennsylvania GPTs-are-GPTs study, and Microsoft's Working with AI paper. Pay, outlook and education come from the US Bureau of Labor Statistics (Employment Projections and Occupational Outlook Handbook, 2024–34 cycle). Every figure on this page is dated and sourced in the Sources section below, and our full method is on the Methodology page.
Sources
- Anthropic, "Labor market impacts of AI" (Massenkoff & McCrory, March 2026) and the Economic Index job-exposure dataset (observed exposure by occupation). Accessed 11 November 2026.
- Eloundou, Manning, Mishkin & Rock, "GPTs are GPTs" (OpenAI / University of Pennsylvania, 2023), occupation-level exposure data (β measure), official dataset. Accessed 11 November 2026.
- Microsoft Research, "Working with AI: Measuring the Applicability of Generative AI to Occupations" (2025), AI applicability scores (CC BY 4.0). Accessed 11 November 2026.
- US Bureau of Labor Statistics, Employment Projections Table 1.2 and Occupational Outlook Handbook (2024–34 cycle): pay, outlook, education. Public domain. Accessed 11 November 2026.
- US Department of Labor, O*NET Database 30.0 (CC BY 4.0): occupation tasks and skills. Accessed 11 November 2026.
Data checked November 2026. Figures are national US figures. Exposure scores describe tasks, not people, and none of the sources above predicts that this occupation will be eliminated.