A number of self-serving biases have recently been explained by asymmetric belief updating under risk which asserts that humans are quick to learn from positive but not negative information. However, risky decisions in real life are often made under ambiguity where only partial information is available about distribution of risks. We demonstrate that under ambiguity, belief updating is not asymmetric but a flexible process of skepticism towards the valence of partially observable facts. When ambiguity size was tractable, belief updating was sensitive to valence: if the information was promising, ambiguity attitude decreased, skeptically balancing the promising prospects of available evidence against the hazards of what might be hidden from the view. Conversely, when the information was disappointing, attitude toward ambiguity increased, cautiously encouraging the participant to be more adventurous than what the available information guaranteed. These results go contradict the predictions from optimistic learning under risk and suggest that belief updating is sensitive to the state of our knowledge and ignorance.