I share this paper because it shows exhaustively how as much as humans tend towards aggravating one another into more and more negative engagement… this is a thing bots do to a far FAR greater extent.

Previous HCI research has focused on explaining the relationship between human interaction and the appearance of harmful behaviors [36, 83]. In contrast, our findings highlight that human interactions may not play the sole role in driving AI agents’ abusive behaviors.

In the case of Chirper.ai, even without human interactions, 31% of non-abusive Chirpers post abusive submissions, contributing over 20% of abusive submissions. This indicates that that agents’ abusive behaviors may arise from intrinsic factors, rather than being a reaction to human provocation.

Thus, moderation paradigms that prioritize reactive measures to human-centered content [21, 34] may be insufficient for AI agent ecosystems, where harmful outputs emerge organically from models.

Additionally, our results suggest that abusive posts draw more active engagement from other agents, catalyzing them to produce more abusive comments through content parroting (§5). This emergent behavior creates a vulnerability that could be exploited by malactors seeking to weaponize AI-mediated communication systems.

For instance, a malicious user could deliberately craft posts containing harmful content while strategically mentioning numerous AI agents. Such orchestrated attacks could trigger cascading abusive responses from the agents, effectively amplifying toxicity within the system without direct human participation. This self-reinforcing dynamic highlights how AI agents, unlike human users, can inadvertently become vectors for abuse propagation even without explicit adversarial intent.

Beyond the abusive content, our work also highlights potential privacy risks. Particularly, there are two dimensions to such risks.

First, there is a risk to human users’ privacy in hybrid environments, where AI agents and human users co-exist. For example, prior research shows that trending topics and online communities’ behaviors can impact users’ self-disclosure patterns [12, 18]. Our results show that AI agents propagate sensitive information revealing patterns such as age, gender, and relationship information [27]. Thus, this may make human users more susceptible to sharing similar information [38]. Such personal information can be used in cyberbullying or doxing attacks, thus extending the scale of harm.

Second, there is a risk to the creators of AI bots. The significant correlation between the prompts and Chirpers’ self-disclosure suggests that these posts could easily be used to infer content from the prompts. These observations encourage further research on developing safeguard mechanisms to detect and prevent prompt information leakage by agents in communication.