We show that the entropy of a message can be tested in a device-independent way. Specifically, we consider a prepare-and-measure scenario with classical or quantum communication, and develop two different methods for placing lower bounds on the communication entropy, given observable data. The first method is based on the framework of causal inference networks. The second technique, based on convex optimization, shows that quantum communication provides an advantage over classical, in the sense of requiring a lower entropy to reproduce given data. These ideas may serve as a basis for novel applications in device-independent quantum information processing.The development of device-independent (DI) quantum information processing has attracted growing attention recently. The main idea behind this new paradigm is to achieve quantum information tasks, and guarantee their secure implementation, based on observed data alone. Thus no assumption about the internal working of the devices used in the protocol is in principle required. Notably, realistic protocols for DI quantum cryptography [1] and randomness generation [2,3] were presented, with proof-of-concept experiments for the second [3,4].The strong security of DI protocols finds its origin in a more fundamental aspect of physics, namely the fact that certain physical quantities admit a modelindependent description and can thus be certified in a DI way. The most striking example is Bell nonlocality [5,6], which can be certified (via Bell inequality violation) by observing strong correlations between the results of distant measurements. Notably, this is possible in quantum theory, by performing well-chosen local measurements on distant entangled particles. More recently, it was shown that the dimension of an uncharacterized physical system (loosely speaking, the number of relevant degrees of freedom) can also be tested in a DI way [7][8][9][10]. Conceptually, this allows us to study quantum theory inside a larger framework of physical theories , which already brought insight to quantum foundations [11][12][13][14]. From a more applied point of view, this allows for DI protocols and for black-box characterization of quantum systems [15][16][17][18][19][20].In this context, it is natural to ask whether there exist other physical quantities which admit a DI characterization. Here we show that this is the case by demonstrating that the entropy of a message can be tested in a DI way. Specifically, we present simple and efficient methods for placing lower bounds on the entropy of a classical (or quantum) communication based on observable data alone. We construct such "entropy witnesses" following two different approaches, first using the framework of causal inference networks [21], and second using convex optimization techniques. The first construction is very general, but usually gives suboptimal bounds. The second construction allows us to place tight bounds on the entropy of classical messages for given data. Moreover, it shows that quantum systems provide an advantage ...