Corruption is pervasive across the world, yet voters keep electing corrupt politicians. One common explanation in the literature is that voters elect corrupt politicians because they do not know they are corrupt. Yet, several studies that randomise the provision of information about corruption find that electoral accountability is weak at best or non-existent at worst. Despite these results, policy-makers still emphasise the importance of transparency and publicity in the fight against corruption. Rather than studying the impact of information about corruption on accountability, we contribute to this debate by taking a different approach, and explore what kind of information allows voters to identify the “bad apples”? Based on a unique dataset which allows us to identify corrupt behaviour unknown to voters, we first employ machine learning algorithms to identify politicians’ political and personal characteristics that are associated with corrupt practices. We subsequently design an experiment that randomises the provision of this information to evaluate what candidate descriptions enable voters to discriminate good from bad politicians. In other words, we explore under what informational conditions voters ultimately pick the right politician. Our study also aims to contribute to the policy push for greater information disclosure about candidates for public office, by refining exactly what within the information leads to better voter choices.