The Hidden Costs of Intelligence

Abstract

As artificial intelligence increases the computational load of models, the tacit environmental costs of predictive models emerge. This study presents a framework for quantifying and optimizing latent costs of intelligence in drug discovery. Benchmarking three methodological families physics-based approaches, machine-learning ensembles and geometric deep learning (GNN) the framework maps trade-offs between predictive accuracy, computational compactness and carbon footprint. Using resistance to kinase inhibitors as an example, GNNs provide high discriminative power (AUC = 0.89) but demand more energy, whereas ensembles offer a better accuracy–sustainability balance, requiring 0.016 kWh per inference. Despite high demands, physics-based methods retain interpretability. A sustainability scorecard combines performance, efficiency and governance indicators into an index (SI), and step-wise escalation workflows relate model complexity to decision criticality, guiding sustainable intelligence in drug discovery.