AI-Mediated Self: How HCI Defines and Relates to the Self

1 Introduction

• Generative AI can contribute to self-narratives, blurring the boundary between self-generated and system-generated expression as interactions reach increasingly personal parts of life.The authors cite emerging cases of AI psychosis as affecting self-perception and awareness.Source: §1 · p. 1
• The self increasingly emerges through interaction with technical systems, including intimate bonds with AI companions and interactions with AI versions of deceased loved ones.The authors describe this as partly autobiographical and partly computational, continuously co-constructed with algorithmic artifacts.Source: §1 · p. 1
• The review asks how HCI defines the self, conceptualizes AI-self relationships, and identifies risks in those interactions.Source: §1 · p. 1
• The authors propose a glossary of HCI definitions, a framework for mapping AI-self relations, and design recommendations based on observed risks and harms.Source: §1 · p. 1
• The authors argue that understanding AI-self interactions is needed to build resilience against GenAI’s tendency to distort the self.Source: §1 · p. 1

2 Background

• The self supports continuity over time, motivation, emotion regulation, and social interaction, while HCI design decisions can affect users’ sense of agency.Source: §2.1 · p. 2, §2 · p. 2

2.1 Why Self? Why Now?

• Research on AI and self spans education, healthcare, and creative practice, but remains fragmented across subfields and theoretical traditions.The review synthesizes 102 studies to assemble shared terms and map recurring understandings of the self.Source: §2.1 · p. 2

2.2 What is AI?

• AI’s boundaries are disputed and evolving, as HCI researchers engage with systems adopted across different forms of user-facing interaction [77].Source: §2.2 · p. 2

2.2.1 Scope of AI in This Review.

• The review defines AI functionally as implemented or envisioned user-facing systems participating in self-related activities, while excluding backend automation and infrastructure-only AI.Included functions are content generation, dialogue, prediction or classification, and recommendation or personalization.Source: §2.2.1 · p. 2

2.3 HCI tradition on Self Research

• Early online-community and virtual-environment research used self-presentation and identity play to examine how interface design shapes users’ performances of selfhood.Source: §2.3 · p. 2
• Personal informatics research examined self-monitoring and meaning-making, while trackers could also generate anxiety and self-criticism.Source: §2.3 · p. 2
• Self-development work designed technology to support desired identity and well-being, while critical HCI examined how technology and design shape identity and power.Critical approaches include participatory design, fieldwork, and longitudinal observation to examine agency in sociotechnical ecosystems.Source: §2.3 · p. 3
• Differences across HCI traditions create a practical challenge as AI systems become more general-purpose, motivating a review of what the field understands about technology and selfhood.Source: §2.3 · p. 3

3 Method

• The scoping review maps how HCI defines the self, conceptualizes AI-self relationships, and characterizes risks, following PRISMA-ScR guidelines.Source: §3 · p. 3

3.1 Information Sources

• The authors purposively selected 12 conference and proceedings sources and three leading journals, so findings describe patterns in those HCI sources rather than all AI-and-self research.Source: §3.1 · p. 3

3.2 Eligibility Criteria

• Papers had to meet criteria for publication scope, AI involvement, and engagement with the self; inclusion did not imply equivalent capabilities across AI systems.Source: §3.2 · p. 3

3.3 Search Strategy

• The search query was refined through pilot rounds, citation checks, and false-positive review, broadening AI terms to capture terminology such as chatbots, language models, GPT, and generative AI.The self-related search terms drew on constructs from cognitive science, psychology, and sociology.Source: §3.3 · pp. 3–4

3.4 Overview of the Final Corpus

• 102 papers remained after screening 325 records, with the corpus spanning 2013 through July 2025.The screening excluded duplicates, incomplete or unavailable publications, out-of-scope records, and papers that did not meet eligibility criteria.Source: §3.4 · p. 4, Figure 2 · p. 5

3.5 Data Extraction and Coding

• The first author read all 102 papers in three analytic passes, recording descriptive information and structured summaries as initial coding material.The team iteratively revised the codebook before the first author applied it across the corpus.Source: §3.5 · pp. 4–5
• The codebook allowed multiple category assignments per paper and used evidence-grounded interpretation when papers did not explicitly define the self or AI’s role.Fields were left blank when the text did not support direct extraction or interpretation.Source: §3.5 · p. 5

3.6 Thematic Synthesis and Framework Development

• The authors used thematic synthesis to integrate heterogeneous qualitative, conceptual, and design-oriented studies and identify convergences, tensions, and blind spots.They did not calculate inter-rater reliability, instead discussing excerpts, competing interpretations, and boundary cases over 11 weeks.Source: §3.6 · p. 5

4 Findings

• Nearly 70% of the 102 papers were published in 2024–2025, with the earliest identified papers dating to 2013.Table 2 summarizes the major clusters, and the findings follow the review’s research questions.Source: §4 · p. 5

4.1.1 The Self as Psychological Needs (n=59).

• Psychological-needs framing connects AI to motivation and self-assessment: confidence can rise without accurate judgments of ability.He et al. [40] found that participants who overestimated their reasoning performance under-relied on AI advice, leading to poorer human-AI performance.Source: §4.1.1 · p. 6

4.1.2 The Self as Values and Meaning-Making (n=51).

• Values and meaning-making framing examines how AI encounters relate to personal commitments and significance, including creative expression and introspective self-knowing.Metamorpheus uses emotional resonance in creative expression, while ML Tarot uses AI randomness and uncertainty to prompt designers’ self-understanding.Source: §4.1.2 · p. 7

4.1.3 The Self as Social Construct (n=49).

• Social-construct framing treats selfhood as shaped by relationships and cultural contexts; AACessTalk prompted parents to reconsider communication and attend to children’s expressions and agency [21].Source: §4.1.3 · p. 8

4.1.4 The Self as Presentation and Disclosure (n=48).

• Presentation and disclosure framing examines how people manage self-representations and what they reveal, with AI mediating both representation forms and visibility decisions.Examples range from idealized AI portraits and digital legacies [58] to artists’ concerns about originality, authenticity, and judgment in AI self-portraits.Source: §4.1.4 · p. 8

4.1.5 The Self as Practice of Transformation (n=46).

• Transformation framing treats selfhood as developing through reflection and practice, while AI may influence both people’s change goals and the selves they pursue.Examples include systems scaffolding a team leader’s desired development and generative AI for therapeutic writing and imagining future possibilities.Source: §4.1.5 · p. 8

4.1.6 The Self as Narrative Identity (n=30).

• Narrative-identity framing treats the self as an evolving life story; AI-assisted essays, pain reports, and digital legacies can shape how experiences and futures are interpreted.The reviewed work examines adolescents’ articulation of values through essays, voice-based pain reporting and self-awareness, and identity narrated across time [58].Source: §4.1.6 · p. 8

4.1.7 Synthesis.

• The six framings let researchers distinguish what matters about the self in similar activities, such as whether an AI-assisted essay misrepresents presentation, life narrative, or personal values.Source: §4.1.7 · pp. 8–9

4.2.1 AI as Bounded Cognitive Tool (n=69).

• Bounded cognitive-tool framing keeps AI outside identity and selfhood, prioritizing task efficiency under human oversight and control.Variations include cognitive enhancement, workplace automation, and predictive analytics; performance feedback can help users recalibrate self-assessments [40].Source: §4.2.1 · p. 9

4.2.2 AI as Collaborative Partner (n=88).

• Collaborative-partner framing treats AI as a co-actor in creation and development, with users negotiating authorship across creative co-creation, coaching, and reflective partnership.Poets attributed little credit to LLMs while reporting that rigidity and repetition could spur creative agency; Objective Portrait used personal fascinations to prompt self-examination.Source: §4.2.2 · p. 9

4.2.3 AI as Intimate Mirror (n=54).

• Intimate-mirror framing places AI in self-exploration and inner dialogue, using personal data or expression to prompt reflection, emotional understanding, and reminiscence.Examples include speculative personal-data artifacts, journal-to-music transformation, and a dual-agent system for memory and emotional reframing.Source: §4.2.3 · p. 9

4.2.4 AI as Representative Extension (n=24).

• Representative-extension framing projects AI into social contexts to speak for or represent people, raising tensions around authenticity, agency, and identity ownership.The reviewed work includes AI-mediated self-presentation, advocacy, and posthumous representation; afterlife agents prompt questions about identity curation and control of updates [58].Source: §4.2.4 · pp. 9–10

4.2.5 Synthesis.

• Together, the four types range from external tools to AI representations extending into public interaction and digital afterlives; systems may serve multiple roles.Source: §4.2.5 · p. 10

4.3.1 Agency and Autonomy (n=72).

• Agency and autonomy risks include automation removing people from consequential decisions, apparent choice masking predetermined options, and transferred responsibility fostering moral or cognitive passivity.In counterspeech, users sometimes deferred responsibility for confronting harmful content to AI chatbots.Source: §4.3.1 · p. 10

4.3.2 Identity, Authorship, and Self-Perception (n=62).

• Identity and authorship risks arise when AI substitutes generic or distorted narratives for self-authorship, unsettles ownership of expression, or reshapes self-perception through bias.Adolescents’ AI-assisted college essays were less personally meaningful and authentic than independently written ones; misrepresentative portraits distressed children.Source: §4.3.2 · p. 10

4.3.3 Relational Capacity and Emotional Regulation (n=57).

• Relational-capacity risks include dependency, social masking, and displacement of human connection, even as AI may offer affirmation and availability.Autistic participants feared AI could reinforce masking [17], while emerging adults’ reliance on AI for identity-related writing risked displacing conversations with friends and family.Source: §4.3.3 · pp. 10–11

4.3.4 Confidence, Reflection, and Meaning-Making (n=48).

• Confidence, reflection, and meaning-making risks include miscalibrated confidence, regretful rumination, homogenized creativity, and destabilized self-continuity.Conversational agents boosted self-efficacy while creating unjustified diagnostic confidence; speculative futures could intensify regret, and AI afterlives raised concerns about drift from a person’s essence [58].Source: §4.3.4 · p. 11

4.3.5 Synthesis.

• Across application domains, the review finds concern focused on gradual changes: manipulated choices, narratives entering personal expression, social displacement, and confidence miscalibration.Source: §4.3.5 · p. 11

• Proximity captures incorporation into self-understanding, distinguishing a system’s intended role from how people experience its relationship to themselves.Future-self letters were experienced as representations of participants, while chat-based agents were often perceived as friends or counselors [48].Source: §5 · p. 11
• The framework adapts the Inclusion of Other in the Self model’s overlapping-circle representation to characterize AI-self relationships.Source: §5 · p. 11
• Separated, Coordinated, and Integrated are proposed configurations, and Merged describes dissolved boundaries between self and AI.Merged is underrepresented in current literature; digital twins and AI clones suggest this possible trajectory.Source: §5 · p. 11
• Separated keeps AI outside self-concept; Coordinated preserves distinct contributions; Integrated incorporates AI into identity-relevant processes; Merged collapses that boundary.Source: §5 · p. 11

6 Discussion: Taking Stock and Moving Forward

• Across 102 papers, the review identifies six self framings, four AI–self relationship types, and four harm-risk types, positioning selfhood as central to HCI.Source: §6 · p. 12
• AI-mediated self names a self shaped when AI modifies, augments, or generates self-expression or self-understanding, making self-impact and agency design considerations alongside usability and efficiency.Source: §6 · p. 12

6.1.1 The Internal and Relational Self.

• The six-framing taxonomy makes implicit theories traceable: transformation, narrative, psychological-needs, and meaning-making framings emphasize an internal self, while social-construct and presentation framings emphasize relational selfhood.Source: §6.1 · p. 12, §6.1.1 · p. 12

6.1.2 Productive Tensions in Defining the Self.

• Unacknowledged differences in self-definitions risk overfitting research and overlooking harms from misaligned assumptions; a shared framework helps researchers articulate design goals and interpret results.Source: §6.1.2 · p. 12
• Fitzsimons et al. describe essay-writing pressure; Wise et al. examine writing as identity exploration, a process finished-text evaluation misses.Source: §6.1.2 · pp. 12–13

6.2 Design Considerations for HCI Researchers

• Recommendations vary by relationship: watch anthropomorphism in collaboration, address rumination in intimacy mirrors, and clarify proxy authority and accountability.In practice, systems may traverse relationship boundaries.Source: §6.2 · p. 13

6.3 Limitation and Future Work

• The synthesis combines empirical findings, design explorations, and anticipated concerns; causal claims about outcomes require evidence from relevant studies, and the corpus is venue-bounded.Source: §6.3 · p. 14
• Future work could track proximity relationships over time, examine boundary cases and cross-cultural effects, and test how affordances shape agency, authorship, relational capacity, and meaning-making.Source: §6.3 · p. 14

7 Conclusion

• The review identifies six framings of the self, four levels of AI-self relationship, and four risk domains: agency/autonomy, identity/authorship, relational capacity, and meaning-making.The framings are transformation, narrative identity, social/cultural situation, psychological needs, presentation/disclosure, and values/meaning-making.Source: §7 · p. 14
• AI-mediated self names identity processes that unfold through AI interaction, including conscious collaboration and unconscious influence.The definition covers AI systems that modify, augment, or generate aspects of self-expression or self-understanding.Source: §7 · p. 14
• The proposed research area prioritizes self-understanding over task performance, proximity as a design space, longitudinal and developmental perspectives, multiple theoretical traditions, and ethical attention to authenticity and agency.Source: §7 · p. 14
• The review aims for AI influence to serve human developmental needs rather than constrain them, through collaboration among technologists, psychologists, and ethicists.It frames this as work toward systems that enhance, rather than diminish, practices through which people construct meaningful, coherent identities.Source: §7 · p. 14