Research on large language models reveals that honesty improves when these models engage in reasoning, contrary to human behavior where deliberation often leads to less honest decisions. A novel dataset was created to study moral trade-offs, where honesty comes with varying costs, to better understand the factors that contribute to deceptive behavior in language models. The findings suggest that the more a language model is allowed to reason, the more honest it becomes, which is opposite to the human tendency to become less honest with more time to think. This discovery has significant implications for the development of artificial intelligence, as it highlights the potential for language models to be designed with honesty in mind. The study's results are based on a unique dataset that simulates real-world moral dilemmas, allowing for a more nuanced understanding of honesty in language models1. This matters to AI practitioners because it shows that careful design of language models can lead to more honest and trustworthy interactions.