The rise of atmospheric CO₂ concentration has boosted global plant photosynthesis for the past decades, contributing to enhanced carbon sequestration by terrestrial ecosystems, and partially mitigating climate change. Yet, the current and future impacts of this phenomenon, known as CO₂ fertilization, remain highly uncertain. In particular, disentangling the effect of rising CO₂ on ecosystem-level photosynthesis from confounding controls (e.g., temperature, humidity, leaf area index) using observational data is challenging. Here we use explainable machine learning and information theory to evaluate CO₂ fertilization effects from globally distributed eddy covariance measurements of gross primary productivity (GPP). We show that incorporating CO₂ concentration in machine learning models substantially improves the estimation of the temporal trend of GPP globally. The resulting machine learning inferred overall CO₂ fertilization effect is within the range of existing estimates from observational proxies and process-based models. The relative attribution of long-term increase in photosynthesis to CO₂ or leaf area index, however, varies depending on the choice of satellite vegetation proxies in the model. Our evaluations of model performance based on causal analysis shows that poorly informed model structures may falsely attribute short-term GPP changes induced by other factors (e.g. drought, heat stress, lack of nutrition) to CO₂ due to overfitting. Our findings show that informing machine learning models with physical knowledge could reduce the uncertainties in differentiating direct and indirect CO₂ fertilization effects on photosynthesis, yet extra caution should be exercised when using machine learning models for causal attribution.