Algorithmic Monoculture in Machine Learning-based Site-Specific Fertilizer Recommendation

Machine learning recommendation systems are commonly evaluated model by model, even when deployment occurs across many decision makers. Fertilizer recommendation creates a consequential setting because a common model can synchronize continuous NPK prescriptions across farms. A formal model is developed in which the recommendation error contains a model-specific common component and a farm-specific component. The expected social loss then contains an adoption concentration term proportional to the Herfindahl-Hirschman Index. Under equal predictive risk, diversified deployment strictly reduces systemic loss whenever the common model error and aggregate externality costs are positive. Under heterogeneous predictive risk, a threshold result establishes conditions under which universal adoption of the individually most accurate model is socially dominated by a mixed model portfolio. Benchmark results across nine model families and five Moroccan cereal systems calibrate the model. A central sensitivity setting with a 10\% private risk regret cap reduces modeled social loss in all five crop systems, with reductions from 0.61\% to 14.19\%, while uniform diversification increases loss in three systems. The results identify model diversity as a potential economic risk control rather than an accuracy objective and motivate ecosystem-level evaluation before concentrated deployment of agricultural recommendation systems.