In this talk I discuss a novel methodology using Bayesian inference to assess how random a sequence of symbols is. Our test assigns a likelihood to all possible models that may have generated a given sequence and then selects the model with the higher likelihood. In this way we are able to infer whether the underlying process was either unbiased or not. In particular, when theoretically compared with Borel normality criterion from algorithmic theory of information and the NIST suite, our method shows a statistical more robust and stricter criterion.