Empathy for pain is often described as automatic. Here, we used implicit measurement and multinomial modeling to formally quantify unintentional empathy for pain: empathy that occurs despite intentions to the contrary. We developed the Pain Identification Task (PIT), a sequential priming task wherein participants judge painfulness of target experiences while trying to avoid influence of prime experiences. Using multinomial modeling, we distinguished three component processes underlying PIT performance: empathy toward target stimuli (Intentional Empathy), empathy toward prime stimuli (Unintentional Empathy), and bias to judge target stimuli as painful (Response Bias). In Experiment 1, imposing a fast (vs. slow) response deadline uniquely reduced Intentional Empathy. In Experiment 2, inducing imagine-self (vs. imagine-other) perspective-taking uniquely increased Unintentional Empathy. In Experiment 3, Intentional and Unintentional Empathy were stronger toward targets with typical (vs. atypical) pain outcomes, suggesting that outcome information matters and that effects on the PIT are not reducible to affective priming. Typicality of pain outcomes more weakly affected performance on an Object Identification Task-wherein target stimuli were merely categorized rather than judged for painfulness-than on the PIT, suggesting that effects on the latter are not reducible to semantic priming. In Experiment 4, Unintentional Empathy was stronger for participants who engaged in costly donation to cancer charities; but this parameter was also high for those who donated to an objectively worse but socially more popular charity, suggesting that overly high empathy may facilitate maladaptive altruism. Theoretical and practical applications of our modeling approach for understanding variation in empathy are discussed.