Debates about algorithmic government have largely focused on transferring decision-making authority from public officials to machines. We ask instead what is lost when AI inference replaces citizens’ voice, and when citizens are willing to delegate voice to AI themselves. We argue that political voice has value beyond the information it provides, but that citizens may be willing to forgo direct expression when participation is costly and their preferences can be reliably inferred. To test these arguments, we propose two preregistered survey experiments involving 4,000 US adults. Study 1 examines whether citizens evaluate government procedures less favourably when their preferences are inferred rather than directly expressed, varying whether public officials or AI process the information while holding accuracy and decision-making authority constant. We expect inference to reduce procedural acceptance by weakening citizens’ sense of being heard, even when their preferences are accurately known. Study 2 is a behavioural experiment where respondents choose whether to express their preferences themselves or allow a predictive system to answer on their behalf. We vary the effort required for direct expression and the system’s predictive accuracy to measure the amount of effort citizens are willing to expend to preserve their voice. We expect delegation to increase with participation costs and predictive accuracy, but to remain less common for political than non-political preferences. Together, the studies contribute a framework for examining the value of political voice in an age of AI, both when it is denied by institutions and when citizens can choose to relinquish it.