Two new arXiv preprints converge on the same worry: people are relying on AI coding assistants without rigorously checking the results. The first study examines software developers and the security of AI-generated code. The authors note that insecure output is well documented, but far less is known about whether and how developers actually evaluate that output before using it.
The second study looks at a different population, researchers who write code for scientific work. Based on 527 free-text responses to a 2025 survey, it maps which tasks researchers hand off to AI assistants and how they decide if the resulting code is correct. The authors find that validation practices vary widely, with little established guidance.
Where the studies differ is focus: one is concerned with security vulnerabilities, the other with scientific correctness. But they agree on the underlying problem—adoption of AI coding tools has run ahead of the practices needed to verify their output. Both papers suggest that trusting AI-generated code without verification is not a viable strategy, and that better validation norms are urgently needed.