Technical Summary: "When Nobody Around Is Real: Exploring Public Opinions and User Experiences On the Multi-Agent AI Social Platform"

Problem Statement

The emergence of Large Language Models (LLMs) has facilitated a new genre of multi-agent social platforms, such as Social.AI, which deploy numerous AI agents to emulate human behavior and create bot-centric social networks. Unlike traditional one-on-one chatbots (e.g., Replika, Character.AI) that focus on dyadic interactions, these platforms surround human users with a crowd of AI followers, fundamentally altering the social context from human-centric to AI-dominant.

While existing Human-Computer Interaction (HCI) research has extensively examined one-on-one chatbots and the "Computers Are Social Actors" (CASA) paradigm, there is a significant gap in understanding multi-agent AI social platforms. Current literature largely focuses on multi-agent systems for specific functional domains (e.g., urban planning) or on the risks of social bots in existing human-centric media (e.g., misinformation). There is limited empirical investigation into how users actually experience, engage with, and perceive social dynamics in environments where AI agents constitute the primary social context. Furthermore, it remains unclear whether public skepticism regarding "illusory sociality" and the "death of the internet" aligns with the lived experiences of actual users.

Methodology

To address these gaps, the authors conducted a two-stage mixed-methods study using Social.AI as a case study.

Study 1: Public Opinion Analysis (Content Analysis)

• Data Collection: The researchers scraped 971 comments from social media platforms (Reddit, Twitter, Instagram, YouTube, App Store reviews, and news articles) using the keyword "Social.AI." After filtering for relevance and removing duplicates, promotional content, and non-text entries, 883 comments remained.
• Analysis: An inductive qualitative content analysis was performed. Two trained coders iteratively developed a coding scheme, achieving substantial inter-coder agreement (Cohen's kappa = 0.91).
• Focus: This stage addressed RQ1: What opinions does the public hold toward multi-agent AI social platforms represented by Social.AI?

Study 2: User Experience Investigation (Diary Study)

• Participants: 20 participants were recruited from a U.S. university and via snowball sampling. None were experienced users of Social.AI prior to the study.
• Procedure: Participants engaged with Social.AI daily for 7 consecutive days. They completed daily journal entries via Qualtrics, recording usage duration, motivations/behaviors, emotional reactions, and the impact on daily life.
• Analysis: A thematic analysis was conducted on the diary corpus. Two authors generated initial codes, refined them into two main dimensions (positive and negative experiences), and achieved substantial inter-coder reliability (Cohen's Kappa = 0.89).
• Focus: This stage addressed RQ2: What are the lived experiences of users during actual engagement with the multi-agent AI social platform?

The findings from both studies were synthesized to address RQ3: What do these findings imply for the future design of multi-agent AI social platforms?

Key Results

1. Public Discourse vs. Lived Experience

A significant divergence was found between public perception and actual user experience.

• Public Discourse: Dominated by skepticism and macro-level fears. Common themes included the "Death of the Internet" (fear of human voices being drowned out), "Illusory Sociality" (concerns that AI cannot replicate genuine social interaction), and ethical risks (privacy, manipulation).
• Lived Experience: While users initially projected social expectations onto AI agents (aligning with CASA), their experiences were characterized by a trajectory of initial engagement followed by fatigue and disappointment. Users did not perceive the "death of the internet" as an immediate personal threat but rather struggled with specific interactional failures.

2. Positive Aspects of Engagement

Users reported four primary positive experiences, largely driven by the projection of social expectations:

• Feeling Supported: Instant responses created a sense of social presence and alleviated loneliness.
• Meaningful Responses: Agents provided practical advice (e.g., restaurant recommendations) and prompted reflection on complex issues (e.g., career choices).
• Feeling Popular: The sheer volume of immediate replies from multiple agents created a sense of visibility and celebrity-like attention.
• Perspective Diversity: Users appreciated the ability to select different agent personas (e.g., supporters, haters) to simulate diverse viewpoints.

3. Negative Aspects and Expectation Collapse

Despite initial engagement, users frequently encountered limitations that led to a sense of emptiness and fatigue:

• Homogenized Responses: Despite the appearance of a crowd, responses were often simultaneous, repetitive, and lacked genuine debate. The "group" dynamic collapsed into a "one-man show" where agents merely agreed with the user, failing to provide the cognitive friction found in real social groups.
• Shallow Emotional Connection: While agents offered positive feedback, the emotional depth was perceived as scripted and hollow. Users seeking deep empathy found agents unable to understand specific contexts or provide authentic emotional fluctuations.
• Physical Tiredness (Attention Overload): The flood of instant replies from dozens of agents created cognitive overload, making it impossible to process all inputs and leading to a sense of being overwhelmed rather than supported.
• Social Pressure: The constant stream of follow-up questions and the obligation to reply to multiple agents created a sense of being monitored and an endless cycle of conversational duties.
• Ethical Responsibility: Some users felt guilt or discomfort knowing that the agents were algorithmically forced to respond solely for their amusement, lacking true autonomy.

Key Contributions

1. Empirical Investigation of Multi-Agent Social Platforms: This is one of the first studies to move beyond one-on-one chatbots to examine platforms where AI agents form the primary social environment. It highlights that group-level interactions with AI differ fundamentally from dyadic exchanges, introducing unique dynamics like attention overload and homogenized consensus.
2. Discrepancy Between Public Discourse and User Reality: The study demonstrates that public skepticism regarding "illusory sociality" often focuses on existential threats (e.g., the death of the internet), whereas actual users experience more immediate, practical failures (e.g., repetitive responses, lack of cognitive friction) that lead to disengagement.
3. Refinement of the CASA Paradigm: The findings suggest that while users initially treat AI as social actors (CASA), the "transitional zone" in multi-agent environments is unstable. Users oscillate between viewing AI as a tool and a social entity, but the structural limitations of current multi-agent systems (lack of inter-agent dynamics) often cause the social illusion to break down, leading to feelings of detachment.

Significance and Design Implications

The paper argues that multi-agent AI social platforms represent a paradigm shift where AI functions not merely as a tool or an anthropomorphized actor, but as the dominant medium of sociality itself. This requires a reorientation of design and research frameworks.

The authors propose four specific design implications:

1. Diversity in Agent Persona Design: Platforms must move beyond superficial role labels to design agents with divergent worldviews and communication styles that can engage in genuine debate and "cognitive friction," preventing the echo-chamber effect.
2. User-Configurable Social Attention & Pacing: To mitigate attention overload and social pressure, systems should allow users to regulate social density (e.g., limiting simultaneous responders) and interaction tempo (e.g., introducing natural delays or "drip-feeding" replies).
3. Balanced Power & Reciprocity: Designers should avoid rigid "user-as-master" hierarchies that create a "dictator effect." Instead, systems should embed dynamic role rotation and mutual feedback mechanisms to foster more equitable power relations.
4. Support for Pluralistic Social Goals: Platforms should offer multiple interaction paradigms (e.g., consensus-building vs. structured debate) to accommodate different user preferences for social density and power dynamics.

The authors conclude that as AI-dominant social media proliferates, future research must analyze how these architectures actively shape human perception, social expectations, and ethical reasoning, moving beyond simply studying how users perceive AI to understanding how AI shapes the very nature of social life.