The future of socially intelligent systems depends on developing abilities to anticipate and empathize with users. Whilst great strides have been made on developing systems for future behavior forecasting that sometimes also claim to do intention estimation, we argue that the predominant state-of-the-art treatment of these problems leads to a significant misunderstanding about this topic. This paper revisits intention estimation, describing the "intention by outcome" problem and how it severely limits a deeper understanding of the nature of the problem. We argue that without a deeper more nuanced understanding of how to develop intention estimation systems, we head into a severely biased world where intentions would only be considered valid by intelligent systems if they came true. Through a case study on estimating unrealized intentions to speak in-the-wild, we highlight open challenges of this largely unexplored topic.