More Code, Less Validation: How Scientists Are Adopting AI Coding Tools

Talk given to the CASS user/developer experience working group by Elle O’Brien of the University of Michigan on Jan 6, 2026. Recording available here. This talk presents findings from a survey of 868 research scientists about their adoption and use of generative AI tools for programming. Adoption is highest among junior researchers and those with less programming experience, with ChatGPT overwhelmingly preferred over targeted developer tools like GitHub Copilot. Scientists reporting the highest perceived productivity gains tend to report less programming experience, lower adoption of development practices like testing and code review, and accepting larger chunks of generated code at once. These patterns raise concerns about research software quality as AI-generated code enters scientific workflows with limited quality control mechanisms.