PaperlayerView exact revision on arXiv

Source-linked AI summary

Rewriting the Self: User Perceptions of AI-Mediated Cognitive Reframing

Yuewen Yang, Olzhas Yessenbayev, Anna R. Van Meter, Tanzeem Choudhury, Daniel A. Adler

arXiv:2610.34647v1cs.HC

TL;DR

AI reframing of personal narratives raises unresolved questions about authenticity, ownership, and identity in therapeutic use. In a mixed-methods study of 28 participants, the authors examined reactions to AI-reframed stressor narratives. Reframing often reduced negative emotions, while participants also reported diminished ownership and current-self authenticity, with perceptions varying by how they understood the AI’s role.

  • Problem

    How AI intervention in therapeutic reframing affects users’ identity, authenticity, and ownership remains underexplored, although it modifies personal narratives and emotional expression.

    Source: §1 · p. 2, §2.3 · pp. 4–5

  • Method

    In a mixed-methods study, 28 participants reviewed GPT-4o-reframed narratives about their stressors and provided survey, think-aloud, and interview data.

    Source: §1 · p. 2, §3.4 · p. 8

  • Results

    AIM-CR often reduced negative emotions and increased aspired-self alignment, while decreasing current-self authenticity and text ownership; participants viewed AI as a “Mirror,” “Another Self,” or “Social Actor.”

    Source: §1 · p. 2

  • Takeaways & Limitations

    Participants’ differing mental models of AI shaped their experience of authenticity, narrative ownership, and reframing, informing design considerations for future AIM-CR tools.

    Source: §5 · p. 17

    Generalizability is constrained by a sample of 28 participants, 27 graduate students from the northeastern U.S., without clinical disorders.

    Source: §5.3 · p. 20

Abstract

from arXiv · show

Individuals are increasingly using AI for psychological support. One specific paradigm, AI-mediated cognitive reframing (AIM-CR), uses AI to automatically reframe users' negative thoughts ("I failed...") into more adaptive perspectives ("I am learning..."). While prior work suggests that AIM-CR tools reduce users' negative emotions, the complexities of AI automatically modifying users' personal narratives remain underexplored. In this work, we investigated these complexities through a mixed-methods study in which 28 participants engaged with an AIM-CR tool that reframed narratives about their stressors. Our findings suggest that participants had mixed perceptions of AIM-CR: although reframing reduced negative emotions for most participants, participants felt less ownership of, and less authenticity in, the reframed narratives. Participants perceived AIM-CR as a validating "mirror," as "another self" that could motivate or alienate them, or as an external social actor. These perceptions shaped therapeutic dynamics. We conclude by discussing design implications for future AIM-CR tools to better preserve user agency in narrative co-construction and more effectively support mental health and well-being.

1 Introduction

AI-mediated cognitive reframing (AIM-CR) may ease negative emotions, but automatically reshaping personal narratives raises questions about authenticity, ownership, and identity. This mixed-methods study examines those tensions and how participants understood the AI’s role.

1 Introduction

  • AIM-CR modifies users’ self-directed text into more constructive viewpoints, drawing on cognitive reframing and narrative-based therapeutic approaches.For example, a self-critical statement about missing a deadline may be reframed as a setback and opportunity to learn [69].Source: §1 · p. 1
  • Prior evaluations report high-quality reframes, reduced negative emotions and emotional intensity, enhanced resilience, and—in some studies—better outcomes than human-written reframes [23] [69] [81].AIM-CR offers a focused alternative to open-ended therapeutic chatbots, which can be derailed across multiple turns [53].Source: §1 · p. 1
  • Because self-narratives support self-regulation, emotional processing, and identity, AI reshaping them raises questions about authenticity and therapeutic outcomes.Authenticity is linked to congruence between internal experiences and expressed selves; unlike traditional therapy, AIM-CR directly changes users’ emotional language.Source: §1 · p. 2
  • Prior AI-mediated communication research reports tensions around authenticity, perceived intrusiveness, agency, and ownership of AI-influenced messages.This study asks how such tensions apply to therapeutic, intrapersonal reframing and users’ identification with the resulting narratives.Source: §1 · p. 2
  • Among 28 participants, AIM-CR often reduced negative emotions and increased alignment with aspired selves, while decreasing current-self authenticity and ownership.Participants understood the AI as a validating “Mirror,” an aspirational but potentially intrusive “Another Self,” or a “Social Actor”; these views shaped therapeutic utility and acceptance.Source: §1 · p. 2
  • The study offers design implications for guiding AI-role perceptions, supporting co-creation and personalization, preserving agency and ownership, and incorporating user context.The authors argue that ethical and effective support for mental well-being requires attention to the psychological dynamics of AI-mediated self-reflection.Source: §1 · p. 2

2 Related Work

Related work connects AI’s effects on language with self-understanding, then describes AIM-CR’s therapeutic promise and the unresolved questions about identity, authenticity, and ownership when AI reframes personal narratives.

2.1 Language and AI-Mediated Communication

  • Language mediates both interpersonal communication and intrapersonal self-understanding, including emotional interpretation and the construction of an evolving identity narrative.AI-mediated communication can bring benefits but also concerns about emotional costs, authenticity, and diminished ownership of AI-generated text [15, 21, 25, 27, 32, 39, 43, 48, 61, 75, 80].Source: §2.1 · p. 3

2.2 Introducing AI Mediation into Therapeutic Reframing

  • AIM-CR systems automatically transform negative thoughts into more constructive perspectives, offering a streamlined, evidence-based alternative to open-ended chatbots [53].Prior evaluations report high-quality reframes, reduced negative emotions, increased commitment and resilience, and possible advantages over human-written reframes [23] [69] [81]. Yet these studies have focused relatively little on perceptions of AI mediation itself.Source: §2.2 · p. 4

2.3 Identity, Authenticity, and Ownership Tensions in AIM-CR

  • How AI-mediated communication’s tensions around identity, authenticity, and authorship apply to therapeutic AIM-CR remains unclear, despite its direct reshaping of users’ emotional expressions and personal narratives.The authors ask how users experience this intervention, perceive reframed narratives’ authenticity, and conceptualize AI’s role; these perceptions may relate to self-disclosure, acceptance, and internalization [14, 23, 35, 41, 51, 53, 65, 70, 82].Source: §2.3 · pp. 4–5
  • The authors use a mixed-methods study to investigate these dynamics and inform responsible design of AI psychotherapeutic interventions.They identify user perceptions of the dynamics as a needed area of investigation for effective design and implementation.Source: §2.3 · p. 5

3 Methods

The study used a single-session, mixed-methods technology probe: 28 participants wrote about a stressor, reviewed a GPT-4o reframe, and reported emotional and narrative perceptions through surveys and qualitative sessions.

3.1 Participants

  • The study recruited 28 participants, 27 of whom were northeastern U.S. university graduate students; none reported current mental-health diagnoses or treatment.Participants were predominantly young adults, and each session lasted approximately one hour.Source: §3.1 · p. 5

3.2 Protocol

  • Participants completed surveys alongside think-aloud sessions and semi-structured interviews to capture quantitative and qualitative data.The study was a single-session technology-probe investigation.Source: §3.2 · p. 5, §3 · p. 5

3.2.1 Measures.

  • Measures covered in-the-moment emotion and intensity, PANAS positive and negative affect, and perceived authenticity and ownership of participants’ texts.The authors selected emotional measures based on prior single-session and self-guided AI-intervention research.Source: §3.2.1 · p. 5

3.2.2 Study Procedure.

  • Participants wrote about a recent stressor, completed a post-writing survey, reviewed an AI reframe, completed a post-reframing survey, and then reflected aloud and interviewed.During think-aloud, they could highlight, strike through, or edit the AI output; interviews probed their reactions and navigation between perspectives.Source: §3.2.2 · p. 6, Figure 1 · p. 6

3.3.1 Prompting AI for Narrative Reframing.

  • GPT-4o was prompted to identify thoughts and distortions, challenge maladaptive patterns, create balanced perspectives, and acknowledge feelings while aligning actions with values and goals.Sequential prompts, few-shot examples [13], and instructions to maintain the user’s writing style guided generation; users saw only the final rewritten narrative [78].Source: §3.3.1 · pp. 6–7

3.3.2 Technology Probe Interface.

  • The web interface let users highlight helpful reframes, strike through unhelpful ones, and edit the generated text, with actions logged for later analysis.Participants accessed the probe remotely.Source: §3.3.2 · p. 7

3.4 Data Collection and Analysis

  • Quantitative analysis compared pre- and post-reframing scores and summarized mean changes and participant improvement or no-change proportions.Meaningful change was defined as more than 0.5 standard deviations from pre-reframing scores; qualitative transcripts were analyzed through reflexive thematic analysis and integrated with quantitative findings.Source: §3.4 · p. 8

3.5 Positionality Statement

  • The author team included computing and information science researchers, a clinical researcher and practicing mental-health clinician, and faculty supervision.The clinician contributed digital mental-health expertise to feedback and revisions.Source: §3.5 · p. 8

4 Findings

The findings describe emotional effects alongside tensions in authenticity and ownership, and examine how participants’ mental models of AI shaped engagement with reframed texts.

4 Findings

  • The findings frame AIM-CR’s effects as including reduced negative emotions and new perspectives, alongside tension between current-self authenticity and ownership versus aspired-self authenticity.The section also examines how participants’ “mirror,” “another self,” and “social actor” models related to engagement with reframed texts.Source: §4 · p. 8
  • Fig. 3 summarizes participants’ post-minus-pre emotional changes: mean PANAS negative affect change was −0.61, while the figure also displays positive affect and emotional-intensity distributions.The figure reports data for 28 participants and marks change categories and mean changes.Source: Figure 3 · p. 9

4.1 Effects of Cognitive Reframing

Cognitive reframing reduced negative affect and emotional intensity while offering participants alternative interpretations, reassurance, and actionable suggestions. Its usefulness depended on suggestions being specific yet flexible and compatible with participants’ circumstances.

4.1.1 Decreased Negative Emotions and Emotional Intensity.

  • Negative affect fell by 0.61 points and emotional intensity by 1.39 points on average; positive affect changed little.Decreases were reported by 60.7% of participants for negative affect and 75.0% for emotional intensity; positive affect increased for 32.1%.Source: §4.1.1 · p. 9

4.1.2 Introducing New Perspectives.

  • AI reframes offered alternative interpretations that helped participants see new possibilities, extend their thinking, and recall perspectives they had overlooked.Participants described the reframes as opening another angle on their experience and prompting more optimistic or empathetic interpretations.Source: §4.1.2 · p. 9

4.1.3 Positivity, Encouragement and Self-Compassion.

  • Reframes could soften self-criticism and offer reassurance, while a community-oriented framing helped one participant feel less alone and see neighbors as allies.Some participants in high distress wanted more direct expressions of comfort.Source: §4.1.3 · p. 10

4.1.4 Reframe Qualities.

  • Actionable, specific suggestions were most useful when grounded in relevant examples, but participants preferred options over fixed directives.Suggestions that ignored personal constraints or situational realities could feel invalidating, underscoring the need for sufficient context.Source: §4.1.4 · p. 10

4.2 Navigating Identity and Authenticity in AI-Reframed Narratives

Participants judged AI-reframed first-person narratives against their circumstances, emotional truth, and self-concept. The reframes could align with an aspired self while feeling less authentic to the current self and less personally owned.

4.2.1 Effects on Authenticity.

  • Aspired-self authenticity rose by 0.61 points on average, while current-self authenticity fell by 1.29 and text ownership by 1.93.Aspired-self authenticity increased for 53.6% of participants; current-self authenticity decreased for 67.9%, and ownership decreased for 78.6%.Source: Figure 4 · p. 11, §4.2.1 · p. 11

4.2.2 Increased Aspired-Self Authenticity.

  • Alignment with an ideal self could validate participants’ desired identity, and first-person suggestions sometimes felt like self-guidance that encouraged action.Participants described reframes as reflecting qualities they hoped to grow into.Source: §4.2.2 · p. 11

4.2.3 Reduced Current-Self Authenticity.

  • Participants sought to preserve their original emotional arc; editing let them soften assertive or overly positive language to better match their present emotional reality.One participant changed “commit” to “try” because the stronger wording did not feel confident to say aloud.Source: §4.2.3 · pp. 12–13
  • Stylistic mismatches, including language that felt robotic or overly optimistic, could undermine authenticity even when participants found the reframes valid.Participants also objected when reframing compressed emotional nuance and smoothed over tensions they wanted to retain.Source: §4.2.3 · pp. 12–13
  • An aspirational portrayal could raise stress when it created unrealistic expectations or suggested steps participants could not take.Participants described excessive positivity as making their darker feelings seem more pronounced.Source: §4.2.3 · p. 13

4.3 Conceptualizing the AI’s Role in Cognitive Reframing

Participants understood AI reframing through different relationships to the generated voice: it could mirror their thoughts, speak as an aspirational self, or act as an external social partner. These models shaped whether reframes felt validating and usable or intrusive and inauthentic.

4.3 Conceptualizing the AI’s Role in Cognitive Reframing

  • Participants perceived the AI as a mirror, another self, or a social actor, shaping whether reframes felt self-generated or like external suggestions.These mental models reflected how participants interpreted the voice and agency behind the reframed text.Source: §4.3 · p. 13

4.3.1 AI as a Mirror.

  • As a mirror, the AI reflected and validated participants’ thoughts, while nuanced wording could make them feel understood and normalize their experiences.Participants also valued reframes that clarified emotions and organized scattered thoughts.Source: §4.3.1 · pp. 13–14
  • First-person reframes often felt easier to internalize and more natural, particularly for participants wary of social judgment.The AI’s first-person tone let some participants reflect without feeling judged as they might when interacting with a person or bot.Source: §4.3.1 · pp. 13–14

4.3.2 AI as Another Self.

  • Some participants experienced the reframe as an aspirational self that could make a more positive perspective feel personally resonant and motivating.P24 described the positive-reframing portion as a conversation with an “inspired self.”Source: §4.3.2 · p. 15, Figure 6 · p. 15
  • When the AI’s assumed first-person voice or advice felt misaligned, participants described intrusion, misrepresentation, and threats to their autonomy over feelings and choices.Participants also objected that the AI lacked the contextual knowledge to authentically speak in their voice.Source: §4.3.2 · pp. 15–16

4.3.3 AI as a Social Actor.

  • Some participants treated first-person reframes as advice from an external entity and evaluated them through social expectations about the AI’s understanding and authority.Advice could feel patronizing, and a participant who viewed the AI as a stranger said its positive feedback mattered less than feedback from trusted sources.Source: §4.3.3 · pp. 16–17

5 Discussion

The discussion frames AI’s role in reframing as a spectrum of perceived self-resemblance, shaped partly by users’ beliefs about AI. It recommends matching the AI’s role to the therapeutic goal while preserving user agency and personal context.

5 Discussion

  • Users’ perceptions of AI as a mirror, another self, or social actor varied with perceived self-resemblance and shaped authenticity, ownership, and the reframing experience.The discussion defines self-resemblance as constructive perspectives that align with users’ values and feel attainable.Source: §5 · p. 17, §5.1.1 · p. 17

5.1.1 The Spectrum of Perceived Self-Resemblance.

  • High self-resemblance made reframes feel like validating mirrors, partial resemblance created an aspirational but potentially alienating self, and low resemblance invited scrutiny of an external actor’s advice.The discussion relates these patterns to self-verification and congruence theories.Source: §5.1.1 · pp. 17–18

5.1.2 Beliefs about AI Technology Influence Engagement.

  • Prior beliefs about AI shaped engagement, while expectations for personalization and broader context could affect whether users trusted its advice.Some participants brought expectations based on tools such as ChatGPT with memory or NotebookLM.Source: §5.1.2 · p. 18

5.2 Designing AIM-CR Tools which Take AI Mediation into Account

  • Designers should guide users toward mental models suited to the therapeutic goal, support control over reframes, and consider user context and prior beliefs where feasible and ethical.The three models offer distinct benefits and risks, motivating deliberate rather than incidental role-setting.Source: §5.2 · p. 18, §5.2.1 · p. 18

5.2.1 Guide Users Towards Appropriate Mental Models.

  • For affirmation and emotional processing, designers can emphasize kind validation and structured reflection rather than introducing transformative perspectives that diverge from users’ present reality.For motivation and growth, the discussion recommends co-creating the aspired self through choices over optimism, wording, and values.Source: §5.2.1 · pp. 18–19

5.2.2 Ensure User Agency in Narrative Co-construction.

  • Because participants varied in preferred reframing style and actively edited outputs, tools should let users choose approaches and accept, reject, or modify suggestions.The discussion also recommends personalizing tone; cited prior work found tone modeling can enhance stylistic resonance and ownership.Source: §5.2.2 · pp. 19–20

5.2.3 Incorporate User Context.

  • Reframes should connect to users’ specific experiences, patterns, and challenges because what soothes or motivates in one situation may feel unhelpful or aggravating in another.The discussion points to prior work using current behaviors or past journals to ground AI support in users’ contexts.Source: §5.2.3 · p. 20

5.3 Limitations

The study’s findings have limits tied to its participant sample, single-session design, and the specific AI model and interface tested.

5.3 Limitations

  • The 28-person sample included 27 graduate students from the northeastern U.S. without clinical disorders, limiting generalizability to broader and clinical populations.The authors call for evaluation in more diverse and clinical settings.Source: §5.3 · p. 20
  • The single-session design does not support conclusions about long-term effects such as sustained engagement or evolving user perceptions.The authors identify longitudinal studies as necessary to track changes with continued use.Source: §5.3 · p. 20
  • Findings were also shaped by GPT-4o and a minimalist interface, so different technologies or more interactive designs could yield different outcomes.The authors call for longitudinal and comparative studies.Source: §5.3 · p. 20

6 Conclusion

The conclusion frames AIM-CR as an intervention that changes self-reflection and users’ experience of their narratives. It describes three perceived AI roles and calls for designs that support beneficial mental models, user control, and trust.

6 Conclusion

  • AIM-CR reshapes self-reflection, with users perceiving the AI as a Mirror, an Aspired Self, or a Social Actor according to alignment with their self-concept.The roles carry distinct potential benefits and risks: validation, motivation alongside possible inadequacy, and novel but potentially untrustworthy perspectives.Source: §6 · p. 20
  • The AI’s voice shapes users’ authenticity, narrative ownership, and therapeutic alliance.Source: §6 · p. 20
  • The conclusion calls for designs that steer toward beneficial mental models, give users meaningful control over narrative co-construction, and build trust through contextual sensitivity.Source: §6 · p. 20

A Writing Prompts for Users

Figure 7 shows the writing prompt used with participants, adapted from the Expressive Writing framework.

A Writing Prompts for Users

  • Figure 7 shows the participant writing prompt, adapted from the Expressive Writing framework [58].Source: Figure 7 · p. 25

B AI Prompts

Table 1 presents the main content of the original prompt used in the technology probe.

B AI Prompts

  • Table 1 shows the main content of the original prompt used in the technology probe.Source: B AI Prompts · p. 25

C Pre-Intervention and Post-Intervention Measures

The study used pre- and post-intervention measures of emotion, affect, authenticity, and ownership. Table 1 describes the prompts used to generate reframes and how outputs were presented.

C Pre-Intervention and Post-Intervention Measures

Related papers

Loading 2610.34647v1…