AI
How software engineers are actually using AI—and what's holding them back
A large empirical study reveals that generative AI tools are deeply embedded in developer workflows, but significant gaps remain between promise and practice.
Generative AI adoption in software engineering is now widespread and deeply integrated into how developers work—yet the reality falls short of the hype. A new empirical study submitted to arXiv examines the current state of GenAI adoption, uncovering both the genuine wins and the persistent friction points that teams face.
Where AI is making the biggest impact#
Developers are leaning on GenAI most heavily for implementation, testing and validation, day-to-day assistance, and maintenance. The tools deliver measurable wins in several areas: teams report faster cycle times, quality gains, better support for knowledge-intensive work, and overall productivity improvements. For many practitioners, AI has become a standard part of the daily workflow rather than an experimental add-on.
The gap between perceived and measured gains#
Despite the optimism, one striking finding emerges: objective measurement of productivity and quality improvements remains rare in practice. While developers feel faster and report better outcomes, few organizations have rigorous data to back up these impressions. This gap between perception and evidence creates blind spots for teams trying to understand whether investments in GenAI are paying off.
Persistent technical and organizational challenges#
The challenges are substantial. Practitioners struggle with unreliable or incorrect outputs, require significant effort in prompt engineering to get useful results, and face a heavy validation burden to catch mistakes. Beyond the technical layer, security and privacy concerns loom large, alongside the risk that teams become overly dependent on AI suggestions without maintaining critical review.
On the institutional side, while many organizations have adopted GenAI tools, adoption is uneven. Companies focus heavily on access and tool deployment but invest far less in training, governance frameworks, and long-term strategy.
What developers expect ahead#
Contrary to dystopian narratives, most practitioners believe AI will . They do harbor moderate concern about job market shifts and the need to acquire new skills, but the prevailing view is one of transformation rather than obsolescence.
