Default Lock-In: A Compute-Economics Account of Cultural Homogenization in AI Writing Tools

Recent empirical work has documented that AI writing assistants homogenize user output toward Western, English-language norms, with measurable productivity asymmetries between Global North and Global South users. Existing accounts treat this as a problem of training data composition, model alignment, or recursive feedback dynamics, and propose remedies focused on data diversification, multilingual fine-tuning, and culturally aware benchmarks. We argue these accounts mistake symptom for cause. Cultural homogenization in deployed AI is best understood as an economic equilibrium sustained by compute asymmetry: under current compute distributions, Western-default foundation models are the cost-minimizing choice for both producers and consumers, and this economic logic, rather than annotation choices or training objectives, is what makes homogenization persistent. We formalize this reframing as default lock-in through three coupled mechanisms: compute concentration, deployment asymmetry, and recursive corpus capture. We argue that purely technical interventions are inadequate because they leave the economic equilibrium untouched. We propose structural interventions, including sovereign compute infrastructure, cost-aware evaluation, and utility-transfer accountability, that target the equilibrium directly. Our core claim is that cultural diversity in AI is not primarily an alignment problem, but a compute-policy problem.