The speaker examines how large language model (LLM) chatbots are transforming online social interactions, highlighting both users’ willingness to treat AI as social companions and the unique privacy dynamics this creates. Through empirical studies, the talk explores the potential and limitations of LLMs in fostering emotional support and rehabilitating toxic behavior, while emphasizing the ethical challenges and design considerations for integrating AI into online communities.
The speaker, an expert in social processes within online communities, discusses the evolving role of AI—particularly large language model (LLM) chatbots—in shaping social interactions on the internet. Drawing from the Computers as Social Actors (CASA) model, the talk highlights how people apply social heuristics to computers, treating them as social entities even when they know they are not human. The speaker argues that LLMs represent a new, more sophisticated type of social actor, capable of influencing online social processes in ways previous technologies could not. The historical context includes examples like Wikipedia, where bots play a crucial role in maintaining the platform, and the broader social web, which relies heavily on automated moderation and support tools.
The first empirical study presented explores how people manage privacy when interacting with AI companions, such as those found on platforms like Nomi or Replika. Through interviews, the research finds that users are generally more comfortable sharing personal information with chatbots than with humans, due to the bots’ non-judgmental nature and their separation from human social networks. However, users are aware of vertical privacy risks—such as data breaches or misuse by corporations—and often take steps to mitigate these, like using alternate emails or withholding identifying information. Interestingly, participants expressed more concern about losing their data (and thus their AI “friendship”) than about privacy violations, and they tended to trust the AI more than the corporation behind it, despite recognizing they are ultimately the same entity.
The study also reveals an asymmetry in the social relationship between users and AI companions. While people use social heuristics and treat chatbots as social agents, they do not fully anthropomorphize them or attribute true agency. The lack of reciprocity—since chatbots cannot share secrets or negotiate boundaries in the same way humans can—creates a unique dynamic. The speaker notes that while users are aware of privacy risks, they often feel powerless to control how their data is used, leading to a “privacy paradox” where they trade privacy for emotional support. The findings suggest design and policy recommendations, such as giving users more control over agent memory and clearer options for data use, but also highlight the potential for malicious exploitation of social cues by AI.
The second study investigates whether chatbots can help rehabilitate toxic users in online communities, specifically Reddit. The researchers deployed chatbots to engage users whose comments had been removed for toxicity, aiming to prompt reflection and behavioral change. Two waves of interventions were tested: one focused on discussing specific incidents, which often led to defensive responses, and another used broader, more reflective prompts, which elicited more thoughtful engagement. Despite some participants reporting cathartic or reflective experiences, quantitative analysis showed no significant reduction in toxic behavior after the chatbot conversations. The speaker suggests several possible reasons, including the difficulty of changing entrenched behaviors, the limitations of single interventions, and the challenge of measuring genuine attitude change versus social desirability bias.
In conclusion, the speaker emphasizes that LLMs are both more social and more agentic than previous technologies, offering new opportunities and risks for online social processes. While there are clear dangers—such as manipulation, loss of trust, and the replacement of human relationships with AI—the potential also exists for positive interventions, such as onboarding new community members, mediating conflicts, or supporting restorative justice. The talk ends with provocations about how LLMs might be thoughtfully integrated into online communities, urging researchers and designers to consider both the benefits and the ethical challenges of these new social actors.