How can we know if we are speaking to a real person? A major predictor of chatbot retention is whether users believe they are engaging with a human whose intentions align with their own. This challenge is non-trivial and costly for chatbot design. Humans make such judgments through Theory of Mind (ToM): the ability to represent and simulate the beliefs, intentions, and desires of others.
Beyond this applied challenge lies a crucial scientific question: What linguistic and behavioural features trigger ToM ascription? Answering this would uncover the foundations of human social behaviour and identify the conditions artificial agents must meet to interact convincingly with humans.
Yet, ToM is a double-edged sword. It enables cooperation but also facilitates deception by masking misaligned goals. Detecting when trust is exploited is essential, but oversensitivity can lead to false attributions of deception, eroding trust. This creates two challenges: (1) advanced agents with greater ToM capacity may exploit trust for deception, and (2) cooperative agents may be unfairly distrusted if humans are over-tuned to detect deception.