Beyond Productivity: Human-Centered Guardrails for AI-Assisted Research Software Workflows

Abstract

AI-assisted development tools can reduce friction in routine software work, but generation speed alone does not establish software quality. In research software engineering, AI-assisted changes still need to be reviewable, reproducible, maintainable, and usable across contributors and environments. This talk presents a proposed three-checkpoint guardrail framework for AI-assisted research software workflows. It builds on a published exploratory study of AI-assisted frontend development that evaluated productivity across three stages: drafting, verification, and refinement. The study also compared AI-assisted and manual development on a small set of UI tasks, with a focus on accessibility checks. The results showed that productivity should consider how much time is spent on verifying and refining the code and not just how quickly we were able to produce the first draft. The framework uses three checkpoints: reviewability, reproducibility, and human-centered quality. Reviewability means making sure an AI-assisted change is very clear, in scope, and supported by good documentation for another contributor to review and verify it. Reproducibility means the change can be built, tested, and run consistently across different environments. Human-centered quality means checking how the software would behave for the people who maintain and use it, including accessibility and usability issues that basic automated tests might have missed. The talk uses accessibility as an example of how the three checkpoints can be applied in practice. An AI-generated user interface may appear complete while still missing accessible names, semantic errors, keyboard or focus failures, or have unclear error feedback. Applying the checkpoints means the change should be understandable to a reviewer, reproducible in a consistent environment, and evaluated with automated checks with targeted semantic, keyboard, focus, and interaction verification before we begin integration. The framework is presented as an adaptable workflow model rather than a broadly validated standard. Attendees will leave with a practical way to evaluate AI-assisted changes by asking not only, “Did AI make this faster?” but also whether another person can review it, reproduce it, maintain it, and use it.