While spreading fake news is an old phenomenon, today social media enables misinformation to instantaneously reach millions of people. Content-based approaches to detect fake news, typically based on automatic text checking, are limited. It is indeed difficult to come up with general checking criteria. Moreover, once the criteria are known to an adversary, the checking can be easily bypassed. On the other hand, it is practically impossible for humans to check every news item, let alone preventing them from becoming viral. We present Credulix, the first content-agnostic system to prevent fake news from going viral. Credulix is implemented as a plugin on top of a social media platform and acts as a vaccine. Human fact-checkers review a small number of popular news items, which helps us estimate the inclination of each user to share fake news. Using the resulting information, we automatically estimate the probability that an unchecked news item is fake. We use a Bayesian approach that resembles Condorcet’s Theorem to compute this probability. We show how this computation can be performed in an incremental, and hence fast manner.