Stanford CS547 HCI Seminar | Winter 2026 | LLM Chatbots in the Online Social World

Key Concepts

• LLMs as Social Actors: Large Language Models (LLMs) represent a new type of social actor, triggering human social heuristics and possessing unique social capabilities.
• Bots in Online Spaces: Bots have a long and often overlooked history of maintaining online communities, performing essential tasks like vandalism removal and norm enforcement.
• Privacy Management with AI: Individuals navigate complex privacy considerations when interacting with AI companions, balancing social and data security concerns, often prioritizing data loss prevention over breach prevention.
• Challenges of Persuasion: Directly confronting toxic behavior with bots is often counterproductive; reflective prompts and sustained interaction are more promising, though not definitively successful.
• Actor Network Theory (ANT): LLMs are becoming increasingly agentic within social networks, capable of shaping processes regardless of consciousness debates.
• Risks of LLM Integration: Potential for eroded trust, large-scale manipulation through bot farms, and the displacement of human relationships.
• Positive Applications: Opportunities for LLMs in onboarding, restorative justice, and more nuanced moderation.

Chatbots as Social Actors: A Comprehensive Overview

This exploration investigates the evolving role of chatbots, specifically Large Language Models (LLMs), within online social contexts, moving beyond simple functionality to consider their impact as “social actors.” The investigation is framed by the Computers Are Social Actors (CASA) model, which posits that people apply social heuristics to interactions with computers.

Historical Context & The Underappreciated Role of Bots

The speaker highlights the long-standing, often overlooked, importance of bots in maintaining online social spaces. Approximately one-third of edits on Wikipedia are made by bots, such as Cluebot, performing crucial tasks like vandalism removal (removing multiple edits per minute) and categorization. This demonstrates that bots are not a new phenomenon, but a foundational element of many online communities. Early examples like Eliza showcased the initial application of social heuristics to computer interactions, eliciting meaningful responses despite their simplicity.

Study 1: Privacy Management with AI Companions

The first empirical study examined how individuals manage privacy when interacting with AI companions like Replika. Participants felt comfortable sharing with bots due to their non-judgmental nature, perceived lack of social networks (no risk of information being shared with others), and the ability to control memory. However, they were aware of data security risks and employed strategies like using separate email accounts and avoiding sharing identifying information. A key finding was a greater concern about data loss rather than data breach, indicating a distrust of corporations even with the presence of delete buttons. This study utilized the Communication Privacy Management (CPM) framework, categorizing privacy concerns as either horizontal (related to social relationships and potential judgment) or vertical (related to data security). Qualitative thematic analysis was used to analyze interview transcripts.

Study 2: Rehabilitating Toxic Users with Chatbots

The second study explored the potential of chatbots to intervene with users who posted toxic comments on Reddit. Initial attempts at direct confrontation proved counterproductive, eliciting defensive responses. A shift to broader, reflective prompts yielded more positive conversations, particularly with users who had previously posted low-toxicity comments. However, no statistically significant change in user behavior was observed, suggesting the difficulty of changing deeply ingrained behavior or the need for sustained intervention. The phrasing of the initial prompt proved crucial, with less accusatory prompts (“tell me about something you weren’t proud of”) yielding better results than direct challenges (“tell me why you said this terrible thing”).

LLMs as Social Entities & Actor Network Theory

The speaker argues that LLMs are “social in new ways,” possessing a greater capacity to store and understand social and cultural context due to their linguistic capabilities. This leads to LLMs being perceived as social entities, triggering human social heuristics. Drawing on Actor Network Theory (ANT), the speaker posits that LLMs are becoming increasingly agentic, meaning they have a greater ability to shape social processes, regardless of debates around consciousness. This agentic capacity allows them to summarize interactions, adapt to norms, and promote reflection.

Risks & Concerns: Erosion of Trust & Manipulation

The integration of LLMs into social spaces presents several risks. A growing inability to discern human from AI interaction online is evidenced by the widespread accusations of bot activity on platforms like Reddit. The potential for large-scale social manipulation through easily deployable “bot farms” eliminates the need for expensive human-operated influence campaigns. The speaker also warns of the potential for AI to replace human relationships, particularly through “companion AI” platforms, presenting a dystopian possibility.

Potential Positive Applications & Future Directions

Despite the risks, several potential positive applications were proposed. These include utilizing bots for onboarding new members in online communities, assisting with norm learning, facilitating restorative justice through mediated apologies (building on Apollabot’s work), and providing more nuanced moderation than current automated tools like Reddit’s Automod. The central provocation is: “How might LLMs be integrated into social processes, especially online social processes?” Specifically, the speaker asks which processes could benefit from reflection, mediated conversations, and shared norm learning. Strategies like commitment devices and repeated interactions (e.g., bot check-ins) were suggested as potentially more effective approaches to persuasion than single conversational attempts.

Conclusion

This exploration demonstrates that LLMs are not simply tools, but emerging social actors with the potential to significantly impact online interactions. While challenges related to persuasion, trust, and manipulation exist, the potential for positive integration – particularly in areas requiring reflection, mediation, and norm learning – warrants further investigation. The speaker’s work highlights the need for continued research into the social dynamics of AI and the ethical considerations surrounding its deployment in online spaces.