The coupling between carbon uptake and water loss through stomata implies that gross primary production (GPP) can be limited by soil water availability through reduced leaf area and/or stomatal conductance. Ecosystem and land-surface models commonly assume that GPP is highest under well-watered conditions and apply a stress function to reduce GPP as soil moisture declines. Optimality considerations, however, suggest that the stress function should depend on climatic aridity: ecosystems adapted to more arid climates should use water more conservatively when soil moisture is high, but maintain unchanged GPP down to a lower critical soil-moisture threshold. We use eddy-covariance flux data to test this hypothesis. We investigate how the light-use efficiency (LUE) of GPP depends on soil moisture across ecosystems representing a wide range of climatic aridity. ‘Well-watered’ GPP is estimated using the sub-daily P model, a first-principles LUE model driven by atmospheric data and remotely sensed vegetation cover. Breakpoint regression is used to relate daily β(θ) (the ratio of flux data–derived GPP to modelled well-watered GPP) to soil moisture estimated via a generic water balance model. The resulting piecewise function describing β(θ) varies with aridity, as hypothesised. Unstressed LUE, even when soil moisture is high, declines with increasing aridity index (AI). So does the critical soil-moisture threshold. Moreover, for any AI value, there exists a soil moisture level at which β(θ) is maximised. This level declines as AI increases. This behaviour is captured by universal non-linear functions relating both unstressed LUE and the critical soil-moisture threshold to AI. Applying these aridity-based functions to predict the site-level response of LUE to soil moisture substantially improves GPP simulation under both water-stressed and unstressed conditions, suggesting a route towards a robust, universal model representation of the effects of low soil moisture on leaf-level photosynthesis.