An AI coding assistant can draft an API endpoint before a developer finishes writing the ticket. That is useful. It still does not answer the question employers and programmers actually care about: will fewer people be needed to build and maintain software?
Predictions about programmers disappearing usually compress three claims into one. AI can produce code. It can make some coding tasks faster. Employers will therefore need fewer developers. Current tools clearly support the first claim, and studies support the second in some settings. The third remains unsettled.
What Productivity Studies Actually Measure
Start with the strongest speed result. A widely cited 2023 experiment recruited 95 professional programmers through Upwork and randomly assigned them to use GitHub Copilot or work without it. Among the 35 completers in each group, the Copilot group finished a JavaScript HTTP-server task in an average of 71 minutes, compared with 161 minutes for the control group. That was a 55.8 percent reduction in completion time. 1
A 90-minute difference matters on a task like this. Even so, the trial covered one well-specified task in one language and did not measure maintainability, security, or performance beyond the supplied test suite. The authors warned against generalizing the result to collaborative projects or other workloads.
Real-world usage data answer a different question. Ziegler and colleagues analyzed Copilot usage and survey responses and found that users accepted 27 percent of shown suggestions. Acceptance correlated modestly with perceived productivity, but the study was observational. It did not establish that accepted code saved time or improved quality. 2
GitHub later reported nearly 30 percent acceptance in product telemetry from Copilot’s first year. Together, those figures show that developers often use the suggestions. They do not tell us whether a project finished sooner, required less review, or produced better code. 3
Now change the setting. In a 2025 randomized trial, 16 experienced open-source developers completed 246 real tasks in mature repositories they had worked on for an average of five years. With early-2025 AI tools available, completion time increased by 19 percent. This result also has a defined scope: experienced maintainers, large familiar codebases, and particular tools at one point in time. 4
These studies point to a range of outcomes rather than one universal productivity effect. The result depends on the task, codebase, developer, tool, and outcome being measured.
Exposure Is Not Job Replacement
Speed on a task is still one step removed from a job. A 2023 paper by Tyna Eloundou and colleagues estimated that about 80 percent of U.S. workers could have at least 10 percent of their tasks affected by large language models, while 19 percent could have at least half of their tasks affected.
In that paper, affected means technically exposed: an LLM or LLM-powered tool might reduce the time required for a task by at least 50 percent while maintaining quality. Exposure could lead to automation, faster work, or a different division of responsibilities. The authors explicitly said their measure does not distinguish augmentation from displacement, and they did not predict adoption timelines or employment effects. 5
The paper found occupation-level associations. Programming and writing tended to correlate with higher exposure, while critical thinking tended to correlate with lower exposure. It did not rank documentation, tests, and code generation above architecture, stakeholder work, and production debugging. It also did not establish a clean junior-versus-senior divide.
Why Employment Does Not Follow Automatically
Labor markets add another layer. When a task becomes cheaper, employers may need less labor for each unit of output. They may also build more, lower prices, or create new work around the technology. Evidence has to establish which effects dominate.
Economists Daron Acemoglu and Pascual Restrepo describe displacement and reinstatement effects. Their framework asks whether automation replaces existing tasks faster than new work appears. It does not turn a productivity gain into a guarantee of employment growth. 6
The familiar ATM example shows why simple stories are risky. U.S. teller employment grew across some selected historical endpoints and later declined. Government series document those counts, but they do not isolate ATM adoption as the cause. In the same way, cheaper code leading to more programmers is a plausible scenario rather than an economic law. 7 8
Forecasts Should Be Labeled as Forecasts
The same discipline applies to claims that solo developers will replace teams, entry-level roles will disappear first, or software demand will expand indefinitely. These are forecasts. The cited studies do not establish them.
Even the seniority story is mixed. The controlled Copilot study found larger gains for less-experienced participants on one task, while the METR trial found experienced maintainers slower on familiar repositories. The settings differ too much to support a universal hierarchy of who benefits and who is displaced.
What Developers and Managers Can Do Now
For developers, treat AI as a tool whose value has to be demonstrated on your work. Count review and correction time, not only generation time. A five-minute draft is not a five-minute saving if understanding and repairing it takes half an hour.
Keep building the skills needed to specify problems, inspect unfamiliar output, test behavior, trace failures, understand systems, and communicate trade-offs. For newer developers, every generated answer should be code you can explain, test, and debug. Current evidence does not prove these skills are immune to automation, but they remain essential for judging what the tool produces.
Managers should measure cycle time, escaped defects, review load, rework, and maintenance rather than lines generated or suggestions accepted. Run pilots on representative work, compare results with a baseline, and record where review or correction consumes the apparent time savings.
AI is already changing how code gets written. The useful question today is concrete: on this team, in this codebase, for this kind of work, does the tool reduce total effort without lowering quality? Measure that result, keep teaching people how to judge the output, and revisit the decision as tools and tasks change. The broader employment consequences remain unresolved.
Primary Sources
- The Impact of AI on Developer Productivity: Evidence from GitHub Copilot
- Productivity Assessment of Neural Code Completion
- The Economic Impact of the AI-Powered Developer Lifecycle and Lessons from GitHub Copilot
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
- GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models
- Automation and New Tasks: How Technology Displaces and Reinstates Labor
- Occupational Employment and Wages May 2010
- Tellers Occupational Outlook Handbook