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

Bots enable online communities at scale

Approximately one-third of Wikipedia edits are performed by bots like Cluebot (an anti-vandalism tool), demonstrating that automated agents have long been essential infrastructure enabling peer production platforms like Wikipedia, Reddit, and GitHub to function.

Evolution beyond traditional CASA models

While the Computers Are Social Actors model established that people apply social heuristics to technology, LLMs represent a new category that is 'more social and more actors' than previous systems, capable of triggering deeper social reciprocity and intimacy cues.

Unique social positioning

Users apply social heuristics to AI companions—developing temporal trust and sharing secrets—while maintaining clear boundaries that these are not human, creating a novel social category distinct from both human relationships and simple tools.

Horizontal vs. vertical privacy mismatch

Users perceive AI interactions through a 'horizontal' privacy lens (social sharing risks like judgment or gossip) and feel safer because bots are non-judgmental and isolated from human social networks, yet the actual risk is 'vertical'—corporate data exploitation and breaches.

The trust paradox

Participants expressed strong trust in their AI companions as confidants while simultaneously distrusting the corporations that own them, despite recognizing these are functionally the same entity, creating a cognitive dissonance around data stewardship.

Data loss anxiety exceeds breach fears

Users demonstrated greater concern about losing their AI companion's memory through corporate failure or technical bugs than about sensitive data being hacked or misused, indicating emotional attachment overrides traditional security risk assessment.

Memory control dilemmas

While users value the ability to edit or delete what AI companions remember about them—treating it as privacy control—this capability introduces dystopian implications of controlling conversational partners' memories in ways impossible in human relationships.

Asymmetrical intimacy vulnerabilities

Unlike human friendships where privacy management involves mutual co-ownership of secrets, AI companions cannot reciprocate with confidential information or make binding promises about data use, creating fundamentally one-sided vulnerabilities ripe for exploitation.

The privacy paradox as helplessness

Users engage in calculated risk acceptance not because they don't value privacy, but because they feel helpless to protect it without sacrificing emotional support, accepting corporate data usage as the unavoidable cost of mental health benefits.

Opportunities for prosocial intervention

Beyond punitive moderation like bans, LLMs offer opportunities for HCI interventions such as toxicity reduction in online communities (e.g., Reddit) and just-in-time warnings about sensitive information disclosure that acknowledge the social nature of AI interactions.