AI Coding Speeds Up Output but Shifts Burden to Senior Developers, Study Finds

Faster code generation is not making software teams move any quicker, according to TechRadar. The publication reports that the industry spent years treating code writing as its main constraint, only to find that removing that bottleneck has not sped up delivery.

Research from Tilburg University in 2025 supports that picture, TechRadar reports. Academics there examined thousands of open-source projects and concluded that AI coding tools lift output while placing a heavier load on experienced engineers. Senior developers devoted considerably more hours to checking machine-written code, and their own production of code fell by close to a fifth.

AI coding agents can now handle tasks that once demanded extra engineering staff, from laying the groundwork of an application to writing tests and pushing implementation forward. A small group of two or three people can test an idea that previously needed a bigger team, deeper specialist knowledge and far more time. According to TechRadar, the bigger shift is that AI raises how much each team attempts, not just how fast it works.

Code volume is not the core problem, the publication argues. Keeping a growing codebase under control is. In enterprise settings, every change still passes through testing, security checks, regulatory requirements and complicated integration work, and those stages cannot keep pace with code generation.

That makes what happens after code is written more important, not less. Organizations are judged on running systems that stay stable, secure and maintainable, which is why managing complexity at scale is described as the defining challenge of the years ahead rather than a shortage of developers.

The unit of innovation is shifting too. Progress no longer depends only on how many people can work on one idea, and team leaders are being pushed to weigh how well groups frame problems, check results and apply human judgement, not simply how quickly they produce code.

TechRadar notes a simple tension at the heart of this: the simpler it becomes to generate code, the simpler it becomes to generate complexity.