Every few months, a new headline announces that AI coding assistants will make software developers obsolete. The claims are dramatic. The timelines are short. And the people making these predictions usually have something to sell.

The actual research on automation and knowledge work tells a different story. It is less exciting but considerably more useful if you are trying to make career decisions based on evidence rather than anxiety.

What the Productivity Studies Actually Found

A 2023 study by Peng and colleagues at Microsoft Research measured the impact of GitHub Copilot on real developer tasks. Developers using the AI assistant completed tasks 55.8 percent faster than those without it. That sounds transformative until you examine what was measured: isolated, well-defined coding tasks given to developers in a controlled setting.

A follow-up study by Ziegler and colleagues found that in real-world usage, developers accepted only about 30 percent of Copilot’s suggestions. The tool was most helpful for boilerplate code, repetitive patterns, and well-documented APIs. For novel architecture decisions, complex debugging, and system design, the AI provided little measurable benefit.

This pattern, automation excelling at routine subtasks while struggling with judgment-intensive work, has appeared in every major automation wave since the introduction of spreadsheets.

The Jevons Paradox in Software

When ATMs were introduced in the 1970s, many predicted the end of bank tellers. By 2010, the number of bank tellers in the United States had actually increased. ATMs made it cheaper to operate a bank branch, so banks opened more branches, which required more tellers for tasks ATMs could not handle: complex transactions, customer relationships, and problem resolution.

Economists call this the Jevons paradox. When a technology makes a resource cheaper to use, total consumption of that resource often increases rather than decreases. Applied to software: if AI makes producing code cheaper, organizations will demand more software, not fewer developers.

McKinsey’s 2024 analysis of enterprise software demand supports this. Companies that adopted AI coding tools increased their total software development spending by an average of 15 percent within 18 months. They did not cut developer headcount. They built more features, automated more processes, and tackled projects that previously failed cost-benefit analysis.

What Is Actually Changing

The research does not support the narrative that AI will replace programmers. But it does support a different and more important claim: AI is changing what programming skills are valuable.

A 2024 study by Eloundou and colleagues at OpenAI estimated that large language models could affect roughly 80 percent of the US workforce by at least 10 percent of their tasks. For software development specifically, the tasks most exposed to automation were code generation from specifications, documentation writing, and test case creation.

The tasks least exposed were system architecture, requirement gathering from stakeholders, debugging production systems under time pressure, and making trade-off decisions with incomplete information.

This maps directly onto the distinction between junior and senior engineering work. The entry-level tasks that traditionally served as training ground for new developers are precisely the tasks that AI handles well. The experienced judgment that senior engineers provide is precisely what AI handles poorly.

The Uncomfortable Skill Shift

If you are a developer whose primary value comes from writing boilerplate code quickly, AI does threaten your role. Not because it will replace you entirely, but because it removes the productivity advantage that was your differentiator.

If your primary value comes from understanding systems, communicating with non-technical stakeholders, debugging complex failures, and making architectural decisions under uncertainty, AI makes you more productive, not less necessary.

The data suggests that the developers who will thrive are those who treat AI as a tool that handles the mechanical work while they focus on the judgment work. The ones who struggle will be those who either refuse to adopt the tools at all or who have built their careers primarily on the mechanical skills that AI now performs adequately.

The Real Timeline

Predictions about AI replacing entire professions have a poor historical track record. A 2019 meta-analysis by Frey and Osborne found that predictions of automation-driven job displacement consistently overestimated the speed and scope of actual displacement by 5 to 15 years. The pattern holds because predictions typically model technical capability in isolation, ignoring organizational inertia, regulatory constraints, integration costs, and the discovery of new tasks that emerge when old ones are automated.

The most likely outcome for programming over the next decade is not replacement but stratification. The floor for what counts as a functional software team will rise. Solo developers will accomplish what small teams do today. Small teams will accomplish what large teams do today. And the total demand for software will continue to expand to fill whatever capacity becomes available.

That is less dramatic than the headlines. But it is what the evidence actually supports.