Generative artificial intelligence has become an everyday tool in software development, with prompting as the primary mode of interaction. Conventional productivity measures capture the outcome but not the cost of reaching it. We call this cost the prompting tax, i.e. the time, effort, and review cost of obtaining working code from an AI assistant. We built an experimental platform combining a code editor, a chat assistant, and automated testing, and ran a user study on three programming tasks. An analysis of 56 tasks solved by 27 participants shows that the tax has a measurable behavioural signature closely tracking perceived effort, that it does not buy higher correctness, that it concentrates on the hardest task, and that it does not ease with practice, while neither confidence nor experience predicts who pays it. We also identify a hidden review tax in the form of deferred verification of proposed code edits, and analyse participants' prompting strategies.
|
