S1_Data.zip

Amidst the growing popularity of AI language models, we study the potential impacts of AI-facilitated communication on trust and trustworthiness. With a laboratory experiment, we compare traditional communication with communication assisted by ChatGPT in a between-subject design. We find that participants with access to ChatGPT make promises more frequently, while their promise-keeping rate diminishes. Overall we do not observe effects on trust and trustworthiness rates. However, we show that in the GPT treatment participants coordinate less often on the trust outcome and that promises no longer serve as a reliable indicator of honesty.