Ivan Soraperra: Effectiveness and Fragility of LLM Apologies

Apologies are a core social mechanism for repairing trust after transgressions. With the increasing availability of large language models (LLMs), individuals can now outsource the writing of apologies, potentially altering both their content and social meaning. In online experiments, this paper studies how LLM-generated apologies differ from human apologies, how they are evaluated, and whether individuals anticipate these differences when choosing how to apologize.

We show that LLM apologies are less variable and more prototypical, relying more heavily on canonical apology strategies that are psychologically costly for humans. When the origin of an apology is undisclosed, LLM and human apologies are similarly effective at eliciting forgiveness. However, a difference emerges with disclosure, highlighting a penalty for machine-generated apologies.

We further examine whether individuals anticipate these dynamics when choosing how to generate an apology. Under disclosure, participants are more likely to write an apology themselves than to use a prewritten LLM-generated apology. Importantly, we observe a similar pattern when the prewritten alternative is human-generated. This suggests that individuals anticipate a penalty for using prewritten apologies per se, rather than a penalty specific to machine origin. Consistent with this interpretation, these effects are fully absorbed once participants’ beliefs about relative effectiveness are controlled for, indicating that apology-generation choices are driven by beliefs rather than disclosure or origin per se.

Overall, we show that LLMs produce apologies that are instrumentally effective but potentially socially fragile when their origin is disclosed. Rather than expanding expressive capacity, LLMs appear to reshape apology and forgiveness outcomes by shifting beliefs about effort and authenticity, thereby altering how actions are interpreted.