Confidence in election results is a central pillar of democracy. However, many new democracies struggle to achieve the democratic ideals. Elections are plagued by a number of irregularities, but neither actors on the ground nor scholars are able to reliably determine whether these irregularities are indicative of fraud, or just reflect growing pains of new democracies. In particular, because existing measures of fraud can only be observed across the entire sample of polling stations, scholars are generally unable to isolate specific fraudulent returns – and thus determine when and where electoral irregularities occur, a crucial question in determining if election results are fraudulent or just problematic. In this paper, we use tools from computer vision and machine learning to identify irregularities on statutory forms for each of over 35,000 polling stations in Kenya’s 2017 presidential election. We show that irregularities on statutory forms correspond to abnormal turnout rates and higher invalid vote rates. Further, we demonstrate that irregularities appear to be concentrated in “stronghold” districts, where one candidate enjoys a preponderance of the vote share.