This paper develops algorithms for improved source selection in social sensing applications that exploit social networks (such as Twitter, Flickr, or other mass dissemination networks) for reporting. The collection point in these applications would simply be authorized to view relevant information from participating clients (either by explicit client-side action or by default such as on Twitter). Social networks, therefore, create unprecedented opportunities for the development of sensing applications, where humans act as sensors or sensor operators, simply by posting their observations or measurements on the shared medium. Resulting social sensing applications, for example, can report traffic speed based on GPS data shared by drivers, or determine damage in the aftermath of a natural disaster based on eye-witness reports. A key problem, when dealing with human sources on social media, is the difficulty in ensuring independence of measurements, making it harder to distinguish fact from rumor. This is because observations posted by one source are available to its neighbors in the social network, who may, in-turn, propagate those observations without verifying their correctness, thus creating correlations and bias. A corner-stone of successful social sensing is therefore to ensure an unbiased sampling of sources that minimizes dependence between them. This paper explores the merits of such diversification. It shows that a diversified sampling is advantageous not only in terms of reducing the number of samples but also in improving our ability to correctly estimate the accuracy of data in social sensing.