A Bell test is a randomized trial that compares experimental observations against the philosophical worldview of local realism , in which the properties of the physical world are independent of our observation of them and no signal travels faster than light. A Bell test requires spatially distributed entanglement, fast and high-efficiency detection and unpredictable measurement settings. Although technology can satisfy the first two of these requirements, the use of physical devices to choose settings in a Bell test involves making assumptions about the physics that one aims to test. Bell himself noted this weakness in using physical setting choices and argued that human 'free will' could be used rigorously to ensure unpredictability in Bell tests . Here we report a set of local-realism tests using human choices, which avoids assumptions about predictability in physics. We recruited about 100,000 human participants to play an online video game that incentivizes fast, sustained input of unpredictable selections and illustrates Bell-test methodology . The participants generated 97,347,490 binary choices, which were directed via a scalable web platform to 12 laboratories on five continents, where 13 experiments tested local realism using photons, single atoms , atomic ensembles and superconducting devices . Over a 12-hour period on 30 November 2016, participants worldwide provided a sustained data flow of over 1,000 bits per second to the experiments, which used different human-generated data to choose each measurement setting. The observed correlations strongly contradict local realism and other realistic positions in bipartite and tripartite scenarios. Project outcomes include closing the 'freedom-of-choice loophole' (the possibility that the setting choices are influenced by 'hidden variables' to correlate with the particle properties ), the utilization of video-game methods for rapid collection of human-generated randomness, and the use of networking techniques for global participation in experimental science.