Do AI coding assistants significantly improve programmer productivity?

How much faster do developers actually work with AI assistants?

The headline numbers are striking. In controlled experiments, developers using GitHub Copilot completed tasks 55.8% faster on average [3]. That's more than halving the time for a typical coding task. At Meta, the multi-line CodeCompose tool saved 17% of all keystrokes — nearly double the 9% saved by single-line suggestions — and multi-line suggestions accounted for 42% of all accepted characters despite being only 16% of displayed suggestions [1]. In Singapore's public sector, developers reported 21-28% faster coding and task completion [5]. These aren't marginal gains; they represent hours saved per week for many developers.

But the picture is more nuanced. A study simulating real development tasks with 27 computer science students found that while AI assistants generally reduced time and improved code quality, experienced participants sometimes saw completion times increase [6]. The tool's suggestions could interrupt their flow or require more mental effort to evaluate and correct. So the biggest time savings go to junior and intermediate developers, while experts may need to adapt their workflow to benefit.

Why do some developers feel less productive despite faster coding?

There's a surprising disconnect: 84% of developers in a longitudinal study reported that AI assistants improved their productivity, yet the proportion reporting worsened developer experience in at least one dimension nearly doubled from 14% to 27% over six months [2]. This 'productivity-experience paradox' means that while tasks get done faster, the mental cost can rise. Developers described a shift from creative coding to what researchers call 'supervisory engineering work' — directing, evaluating, and correcting AI output [2]. This can erode flow state and increase cognitive load, even as feedback loops improve.

The same study found that 82% of developers spent less time writing code, but they didn't necessarily feel better about their work [2]. At IBM, surveys of 669 users showed that while net productivity increased, not everyone experienced the benefit equally [4]. Some developers worried about code ownership and responsibility for AI-generated code [4]. The lesson: productivity metrics like keystrokes saved don't capture the full human experience of programming.

Who benefits most, and when do AI assistants fall short?

Junior developers and learners gain the most. A coding bootcamp case study found that GitHub Copilot significantly improved learning efficiency, code quality, and developer confidence for beginners [9]. In Singapore's public sector, junior developers experienced the largest efficiency gains and reduced coding times [5]. The tool acts as an always-available tutor and accelerator, helping novices produce working code faster and learn patterns.

However, AI assistants struggle with tasks requiring deep logic, unusual edge cases, or strict output formatting. A study of three popular tools found that 'edited line completion' was the most accepted suggestion type, while 'comments completion' and 'string completion' had the lowest acceptance rates [6]. The main reasons for rejecting suggestions were flawed logic, mismatched output formats, and inconsistent code style [6]. Security and accuracy remain concerns: AI-generated code can contain vulnerabilities or subtle bugs [7][8]. The consensus across multiple papers is that these tools are best used as supportive aids, not replacements for human judgment [7][8][10].

Sources used in this answer

Related research questions