Towards Individually Adapted AI-Generated Code Documentation and Its Effects on Developer Understanding

Sammanfattning

Large language models (LLMs) are increasingly used to automatically generate software documentation, yet most existing approaches focus on general documentation quality rather than how documentation supports different types of readers. This thesis investigates whether explicit audience adaptation in LLM-generated software documentation produces measurable differences in comprehension, cognitive load, and perceived usefulness compared to standard audience-implicit documentation. The study uses Design Science Research as the overarching methodology, in which a documentation generation artefact was developed and then evaluated through a controlled between-subjects experiment. The artefact generates documentation from source code using prompt-based audience descriptions. The evaluation ran on a purpose-built web platform, in which participants completed a Python programming task using either standard or adapted documentation. Data was collected from Swedish higher-education students through task performance measures, NASA-TLX, the Paas mental effort scale, custom Likert items, behavioural interaction logging, and open-ended feedback. The results showed that adapted documentation was perceived as significantly more relevant than standard documentation. Exploratory subgroup analyses suggested that less experienced participants benefited most from the adapted documentation, with higher task correctness and lower perceived workload. More experienced participants showed smaller or reversed effects on some task measures, consistent with the Expertise Reversal Effect within Cognitive Load Theory. These findings suggest that audience-aware documentation generation may improve the effectiveness of AI-generated developer documentation, with adaptation effectiveness depending on the match between target persona and actual reader.