Is AI Actually Replacing Software Jobs?

The phrase “AI will replace programmers” is easy to say because it compresses several different questions into one sentence. Can AI write code? Yes. Can companies automate some programming tasks? Yes. Does that automatically mean fewer software developers overall? The public employment data does not currently show that.

That distinction matters because “programmer,” “software developer,” “software engineer,” “QA analyst,” “web developer,” and “AI specialist” are not interchangeable labour-market categories. They overlap in skills, but they describe different bundles of tasks and different employer needs.

When we separate them, a much more interesting pattern appears: the occupation most closely associated with writing and maintaining code from detailed specifications is projected to shrink, while the broader occupation responsible for designing, building and maintaining software systems is projected to grow.

One headline hides two different job markets

The latest BLS 2025–35 projections put software developers at 1.718 million jobs in 2025 and 1.893 million in 2035 — an increase of about 174,700 jobs, or 10.2%. Computer programmers, by contrast, are projected to fall from about 110,800 to 102,700, a decline of 8,100 jobs, or 7.3%.

This is not evidence that one job title is “good” and another is “bad.” It is evidence that the labour market is reorganising around different task bundles. BLS says repetitive programming work is increasingly automated and that some higher-skilled programming tasks are shifting to software developers.

Software developers are still projected to grow

BLS had already examined this issue directly in a 2025 case study on occupations exposed to generative AI. Using its 2023–33 projection cycle, the agency projected software-developer employment to rise 17.9%, from 1.692 million to 1.996 million jobs. BLS explicitly noted that AI could raise developer productivity while also creating demand for people who build AI-enabled products and maintain AI systems.

The newer 2025–35 projection is less aggressive — 10.2% rather than 17.9% — but it is still positive and still well above the projected 3.5% growth for all U.S. occupations combined.

There is also a structural reason the software-developer category can grow while code generation becomes cheaper. Software demand itself can expand. If lower-cost coding lets companies build more products, automate more workflows and embed software in more industries, total demand for people who design systems may continue to rise even when each individual developer becomes more productive.

Computer programmers are a different story

The computer-programmer projection is the clearest evidence that AI and automation can reduce demand for a specific type of software work. BLS says companies are expected to automate repetitive programming tasks and move some higher-skilled work into adjacent occupations such as software development.

That does not mean “coding is dying.” Coding is a task, not one occupational category. Software developers code. Data scientists code. Security engineers code. Researchers code. Analysts increasingly code. The labour-market question is who owns the surrounding product, architecture, testing, deployment and business decisions.

This is why a count of “programmer” jobs cannot be used as a proxy for all software work. The more plausible transition is from narrower execution roles toward broader development roles that combine coding with system design, product context, integration and judgement.

Developer activity is not collapsing

Employment data tells us what employers are doing. GitHub tells us something different: how much software-building activity is happening in public and open-source ecosystems.

In 2024, GitHub reported a 98% year-over-year increase in public generative-AI projects and a 59% increase in contributions to those projects. Python became the most-used language on GitHub by overall activity, with the company linking the change partly to the rise of AI, data science and machine learning.

GitHub's open-source survey also found 73% of respondents said they used AI tools for coding or documentation. That is not an employment statistic, and it should not be treated as one. But it is strong evidence that software work is absorbing AI tools at scale rather than simply vanishing.

The more interesting question may therefore be: what does a developer do when code generation becomes a standard capability? The data suggests more time may move toward reviewing, integrating, testing, designing and deciding what should be built in the first place.

AI demand is real — and geographically concentrated

Stanford's 2025 AI Index uses Lightcast job-posting data to show where AI hiring demand is concentrated in the United States. California accounted for 15.7% of U.S. AI job postings in 2024, followed by Texas at 8.8%, New York at 5.8%, Virginia at 4.7% and Washington at 4.7%.

This distribution matters because national averages can hide local labour markets. A developer in Seattle, San Francisco or New York is exposed to a different mix of AI employers than someone in a region where AI hiring is sparse.

The World Economic Forum's 2025 employer survey points in a similar direction from a different dataset: AI and machine-learning specialists are among the roles employers expect to grow fastest through 2030, and software/application developers also appear among the fastest-growing roles.

Why tech layoffs do not answer the AI question by themselves

Another reason the public debate becomes confusing is that layoffs are often treated as proof of AI displacement. A layoff announcement is real evidence about one company's staffing decision, but it does not tell us why the decision happened or whether total employment in an occupation is rising or falling.

Technology companies expanded very quickly during the low-interest-rate and pandemic periods. Later reductions can reflect a correction in that hiring, weaker product demand, reorganisations, mergers, outsourcing, profitability targets, changes in capital costs or AI-enabled productivity. Several of those forces can happen at the same time.

That is why this page does not count layoff headlines and label the total “AI job losses.” Doing so would assign a cause that most layoff announcements do not establish. Instead, we look for datasets that define the occupation and measure employment or expected employment consistently.

There is a second complication: a company can reduce headcount while an occupation still grows nationally. Workers can move between employers, startups can expand while large firms contract, and software roles can spread into industries that are not usually thought of as technology companies. Banks, retailers, manufacturers, logistics businesses and healthcare organisations all hire software workers.

The entry-level market may be a different problem

Occupational growth does not tell us who gets the new jobs. A market can add software-developer positions overall while becoming harder for junior candidates if AI tools reduce the amount of simple work traditionally given to early-career employees.

The public datasets used in this analysis are not detailed enough to establish that effect. BLS projections describe occupations, not the internal seniority mix of every development team. GitHub activity can show participation, but it cannot tell us whether a contributor has been hired into a paid role.

That means claims such as “AI has eliminated junior developers” require a different evidence base — ideally job-posting data segmented by required experience, hiring outcomes and longitudinal workforce data. We have not used a weaker proxy to manufacture an answer.

This distinction is important for readers because the practical experience of looking for a first software job can deteriorate even while the long-run occupation remains a growth category. Aggregate employment and entry-level opportunity are related, but they are not the same metric.

AI can reduce coding time without reducing software output

Productivity changes create another counterintuitive possibility. Suppose a development team can produce a feature in six days instead of ten because code generation, test generation and documentation are faster. One possible outcome is fewer developers. Another is the same number of developers shipping more features. A third is a company attempting projects that previously would not have been economical.

The labour result depends on how much demand expands relative to productivity. Economists have seen versions of this pattern with earlier technologies: automation can remove particular tasks while lowering production costs enough to increase overall output. Software is unusually exposed to this mechanism because additional software can be distributed at very low marginal cost and because almost every industry is still digitising processes.

We therefore should not infer employment directly from how impressive a coding model looks in a benchmark. A model capability test answers “can the system perform this task?” Employment data answers “how many people do organisations employ to accomplish the whole bundle of work?” The gap between those two questions is where most of the uncertainty lives.

What the data actually says

That conclusion is less dramatic than “AI will replace developers,” but it is also more useful. It suggests three simultaneous changes:

1. Task automation: AI can remove or accelerate repetitive coding, documentation, test generation and debugging work.
2. Role recomposition: jobs may require more system design, product understanding, AI integration, review and responsibility for outcomes.
3. Demand expansion: cheaper software production can create more software projects and more demand in AI-heavy sectors, offsetting some productivity-driven labour reductions.

The relative strength of those forces will determine employment. We do not yet have enough historical data to say which force will dominate across every company, country and skill level.

What we would watch next

The strongest future evidence would combine four measures over the same period: actual employment by software occupation, the seniority mix of job postings, compensation, and the share of developer tasks performed with AI assistance. If employment remains positive while junior postings fall and output rises, the story would look very different from a broad collapse in software work.

We would also watch the boundary between “computer programmer” and “software developer.” If organisations increasingly redesign jobs rather than simply removing them, occupational titles and task definitions may shift faster than the total amount of software work.

Finally, geographic concentration matters. AI-heavy labour markets can expand while other regions stagnate. A national average should therefore be paired with state or metro-level data before making claims about an individual worker's local opportunity.

What this data does not prove

One of the biggest weaknesses in AI-job discussions is that evidence with different meanings gets mixed together: layoffs are treated as employment trends, GitHub activity as hiring, job postings as jobs filled, and survey expectations as observed outcomes. KalQLater keeps those measures separate.

Methodology

We treated “software jobs” as a group of distinct occupations rather than one category. Employment projections come from the U.S. Bureau of Labor Statistics. Developer activity comes from GitHub's Octoverse. AI-posting geography comes from Stanford's AI Index using Lightcast data. Employer expectations are used only as contextual evidence, not as observed labour-market outcomes.

Frequently asked questions

Sources