AI
New framework helps engineers write clearer AI explainability requirements faster
Researchers have developed a guideline-based workflow and tool that cuts time spent drafting explainability requirements for AI systems while maintaining quality comparable to manually written ones.
Making AI systems transparent and trustworthy requires clear specifications—but writing explainability requirements is often haphazard and can leave teams with vague or incomplete directives. A new approach from the software engineering research community aims to solve this by combining structured guidelines with lightweight AI tooling.
The challenge with current practice#
As AI becomes embedded in critical software systems, regulators and users increasingly demand transparency about how these systems work. Requirements engineers must translate that demand into actionable specifications, yet they typically operate without systematic guidance. Letting language models draft requirements without oversight can produce statements that contradict each other, leave key details unspecified, or miss practical constraints—making them harder for developers to implement.
A structured guideline approach#
The research team first conducted a literature review and interviews with developers to identify what makes a good explainability requirement. They then surveyed practitioners to narrow the field down to ten core quality properties that matter most. These include dimensions like clarity, measurability, and implementation feasibility. The team packaged these properties into a concise guideline with step-by-step instructions for formulating each one.
Tool support and validation#
Researchers built a web-based tool that walks requirements engineers through an iterative workflow: draft a requirement, run automated checks against the ten quality properties, and revise based on feedback. Two empirical studies tested whether this approach actually helps:
- Time savings: In a controlled task with six requirements engineers, tool-supported drafting took 23.5% less time than writing from scratch (statistically significant with p = 0.049).
- Quality outcomes: A separate study with eighteen developers found no significant difference between tool-assisted and manually written requirements in terms of how well they could be implemented or their perceived quality—with a slight descriptive edge toward the tool versions.
What this means for practice#
The findings suggest that a prioritized quality guideline combined with AI assistance can reduce effort without compromising the substance of what engineers deliver. The approach avoids heavy-handed automation that might oversimplify requirements, instead using guidelines and lightweight checks to scaffold human decision-making.
