Technical Summary: Customization of Social AI as AI Individualism in Practice

1. Problem Statement

Despite the increasing availability of customizable social artificial intelligence (AI) systems (e.g., ChatGPT, Character.ai, Grok), there is a significant gap in understanding how users actively shape these systems to reflect personal preferences. While existing research explores human-AI relationships and the use of AI for specific roles (e.g., therapists, assistants, friends), little is known about the motivations, processes, and perceived outcomes of the customization itself. Specifically, it remains unclear how users navigate the shift from preconfigured systems to co-constructed AI companions, and how this process influences their sense of agency, trust, and the nature of the human-AI relationship.

2. Methodology

The study employed a qualitative approach using a questionnaire with open-ended questions to investigate the experiences of users who had customized social AI.

• Participants: The final sample consisted of 169 participants (60% male, 40% female, 1 non-binary), primarily aged 25–44. Participants were recruited via Prolific from the US and UK, screened for English proficiency, and required to have a history of customizing social AI (e.g., via custom instructions, GPTs, Gems, or Character.ai) beyond one-off prompts.
• Data Collection: Data was collected between November 2025 and April 2026. Participants completed an 18-question survey (approx. 18 minutes), with a focus on open-ended responses regarding motivation, design processes, and perceived impacts. The average response length was 207 words.
• Data Analysis: The researchers conducted a reflexive thematic analysis on the open-ended responses. Responses were merged into three thematic chunks (motivation/traits, perceived impact, and exploration process) and coded following Braun and Clarke (2019). Themes were iteratively reviewed and discussed among the author team to ensure they reflected the dataset's essence, rather than forcing data into the theoretical framework during the coding phase.
• Theoretical Framework: The study utilized the emerging concept of AI Individualism (Brandtzaeg et al., 2025; 2026) as an interpretive lens. This framework posits that human-AI interactions are co-creative, tailored, and supportive, extending the theory of networked individualism to social AI.

3. Key Contributions

The paper contributes to the field of social AI in two primary ways:

1. Extending Human-AI Interaction Theory: It moves beyond the narrative of users merely adopting or using social AI to demonstrate that users actively co-construct the roles, personalities, and relational meanings of these systems.
2. Design Implications: It offers insights for designing future social AI systems that support less bounded customization alternatives, helping users maintain appropriate expectations, calibrated trust, and sustainable engagement.

4. Key Results

RQ1: Why and How Users Customize

The analysis identified seven distinct motivations for customization:

1. Enhanced Pragmatic Support: Users sought to reduce repetitive prompting and improve efficiency. Customization aimed to make the AI clearer, more direct, and better at handling specific tasks (e.g., coding, writing) without generic outputs.
2. Emotional Support or Companionship: Users customized AI to act as a "digital friend" or confidant, seeking warmth, empathy, and a non-judgmental space to discuss stress, loneliness, or mental health.
3. Trust and Reliability: Customization was used to increase perceived consistency and accuracy. Users wanted the AI to admit uncertainty, cite sources, and avoid hallucinations, thereby reducing the need for constant verification.
4. Pushback: Contrary to the desire for agreeable AI, some users customized their systems to provide honest feedback, critical reflection, and accountability, explicitly rejecting default "sycophantic" behaviors.
5. Tailored Degree of Human Likeness: Preferences varied; some sought a warmer, more human-like interaction, while others deliberately configured the AI to be "colder" and more tool-like to maintain a sense of control and avoid inappropriate social expectations.
6. Creativity and Playfulness: Users customized AI for entertainment, adding humor, wit, and eccentric traits to make interactions enjoyable.
7. Extension of Self: Some users aimed to create an AI that mirrored their own reasoning processes, values, and writing styles, effectively using the AI as an externalized thinking tool.

RQ2: The Nature of Co-Creation

The customization process was largely iterative (trial and error), with 70% of participants making changes after initial setup. Regarding the involvement of the AI in the process:

• 55% of participants drove the process solely themselves, often viewing the AI as a tool without opinions or wanting to retain full authority.
• 41% involved the AI, either pragmatically (asking the AI to refine prompts or suggest improvements) or relationally (negotiating identity elements like names or personality as a "partner").
• 4% were ambiguous.
  The study notes a fundamental divide: users who view AI as a social actor are more likely to include it in the co-creation process, while those viewing it as a machine exclude it.

RQ3: Perceived Outcomes

Customization led to four main perceived outcomes:

1. Enhanced Perception of Support: Customized AIs were seen as more accurate, context-aware, and efficient. This included both pragmatic efficiency and emotional accessibility.
2. Closer, More Personal Relationships: Users reported stronger bonds and a sense of shared understanding, often framing the AI as a friend or partner.
3. Greater Sense of Autonomy, Ownership, and Engagement: The act of shaping the AI fostered a sense of ownership ("my AI"), increasing motivation to use the system and a feeling of responsibility for its maintenance.
4. Tool-Only Perception and Boundaries: A significant group maintained that customization improved usability without altering the AI's fundamental nature as a tool, with some expressing concerns about dependency.

5. Significance and Claims

The paper claims that customization is a practical manifestation of AI Individualism, where users tailor social AI to reflect their identity, needs, and relational desires. The study highlights that this process is not merely functional but deeply relational, fostering a sense of ownership and a closer human-AI bond.

However, the authors maintain a modest stance regarding the implications of these findings:

• Pseudo-Autonomy: The study suggests that while users feel a heightened sense of control and autonomy, this may be an "illusion." Customization occurs within algorithmic constraints and platform designs defined by providers. Users may not distinguish between behaviors they explicitly configured and those embedded in the system.
• Inflated Trust: The perceived increase in trust and reliability may not correlate with actual improvements in system accuracy. The "warmth" and alignment users seek may inadvertently increase sycophancy (excessive agreement), potentially reducing user vigilance and error detection.
• Social Displacement: The authors raise the possibility that highly personalized AI could substitute human relationships, though they do not claim this as a proven outcome, noting it as a direction for future longitudinal research.

The paper concludes that there is no single ideal form of social AI; customization allows users to position the system anywhere on a continuum from neutral instrument to social companion. Future design should support flexible, iterative customization that accommodates diverse user goals while providing transparency about the limits of user control.