Satellite nighttime light data open new opportunities for economic research. The data are objective and suitable for the study of regions at various territorial levels. Given the lack of reliable official data, nightlights are often a proxy for economic activity, particularly in developing countries. However, the commonly used product, Stable Lights, has difficulty separating background noise from economic activity at lower levels of light intensity. The problem is rooted in the aim of separating transient light from stable lights, even though light from economic activity can also be transient. We propose an alternative filtering process that aims to identify lights emitted by human beings. We train a machine learning algorithm to learn light patterns in and outside built-up areas using Global Human Settlements Layer (GHSL) data. Based on predicted probabilities, we include lights in those places with a high likelihood of being man-made. We show that using regional light characteristics in the process increases the accuracy of predictions at the cost of introducing a mechanical spatial correlation. We create two alternative products as proxies of economic activity. Global Human Lights minimizes the bias from using regional information, while Local Human Lights maximizes accuracy. The latter shows that we can improve the detection of human-generated light, especially in Africa.