X-pressing Intimacy: Mental Health Disclosure After Twitter Acquisition

1 Introduction

• The study examines how mental health expression and disclosure intimacy changed around Twitter’s acquisition, guided by Social Penetration Theory.It asks how symptomatic and psycholinguistic expressions changed and how disclosure intimacy changed across the transition.Source: §1 · p. 2
• The analysis compares pre- and post-acquisition posts across five mental health topics, using comparison tweets to distinguish mental-health language shifts from broader Twitter changes.Intimacy is measured on a three-level scale from less personal references to highly intimate accounts of lived experiences.Source: §1 · p. 2
• Only the relative shift from moderate- to low-intimacy BPD disclosures remained statistically significant against the control corpus.Anxiety, depression, and BPD showed shifts toward low-intimacy disclosures, while SI-related posts showed more high-intimacy disclosures; the other condition patterns were not significant relative to control.Source: §1 · p. 2
• After the acquisition, mental health posts showed declines in affective and social language alongside increases in first-person singular pronouns and future-focused language.Declines included anxiety and sadness; intimacy patterns varied by condition.Source: §1 · p. 2
• The study combines psycholinguistic analysis with disclosure-intimacy measures to examine sensitive self-disclosure during a major platform transition.The authors connect this approach to designing online environments that support sensitive self-disclosure and mental health support-seeking.Source: §1 · p. 2

2.1 Mental Health Self-Disclosure on Social Media

• Self-disclosure depth is shaped by perceived rewards and costs, audience expectations, community norms, platform affordances, and moderation practices.Prior work also relates audience engagement to subsequent changes in disclosure intimacy.Source: §2.1 · pp. 2–3

2.2 Mental Health, Psycholinguistics, and Social Media

• LIWC and social-media language research support population-level study of mental-health expression, while this work adds disclosure intimacy to distinguish language from sharing depth.Prior studies have examined language around mental-health conditions and major events; construct validity and methodological limitations warrant caution.Source: §2.2 · p. 3

2.3 Platform Governance and Changes in Online Participation

• Research on Twitter’s acquisition has examined content, participation, sentiment, and migration, but less is known about sensitive first-person expression around the transition.Schizophrenia stigmatization research differs from studying disclosure of one’s own mental-health experiences.Source: §2.3 · p. 4

3 Overview of Study Design and Data

• The design divides Twitter posts at the October 27, 2022 acquisition and measures both LIWC psycholinguistic attributes and low-, moderate-, and high-intimacy disclosure.The observation window runs from January 2018 to February 2023 and includes five mental-health conditions plus general Twitter posts.Source: §3 · p. 4

3.1 Twitter Data

• The data comprise English-language posts explicitly referring to users’ own depression, bipolar disorder, anxiety, BPD, or suicidal-ideation-related experiences.Searches used 40 phrases, including current and past-tense wording for suicidal ideation, desire, intent, or attempts; matching phrases do not verify clinical diagnoses or literal disclosure.Source: §3.1 · p. 4

3.2 Compiling Treatment and Control Data

• The five treatment samples may overlap, while the independent control sample contains 1,000 Twitter posts per week over the same observation period.The control provides a temporal reference for whether linguistic changes also appeared in general Twitter posts.Source: §3.2 · p. 4

4.1.1 Psycholinguistic Analysis

• LIWC2015 scores each tweet by the proportion of words in linguistic and psychological categories across treated and control corpora.Categories span affect, cognition, perception, interpersonal focus, temporal references, grammar, biological and social concerns, and informal language.Source: §4.1.1 · pp. 4–5

4.1.2 Measuring Psycholinguistic Changes

• For each category, the analysis calculates pre/post mean scores and relative percentage changes, then uses within-group t-tests with Bonferroni correction.The method reports means, relative changes, and adjusted significance.Source: §4.1.2 · p. 5

4.2.1 Manual Annotation

• Manual labels define low, moderate, and high intimacy by how much personal context, feelings, symptoms, or private experience a post discloses.A diagnosis or SI reference alone does not determine intimacy; two annotators labeled posts and a third resolved disagreements, with Cohen’s κ = 0.7323.Source: §4.2.1 · p. 5

4.2.2 Automated Classification

• Qwen3-8B classifies intimacy using the codebook and retrieved labeled examples; on 40 held-out posts it achieved 90.0% accuracy and a macro-F1 score of 0.897.All four errors were low-intimacy posts classified as moderate; moderate- and high-intimacy posts were correctly classified.Source: §4.2.2 · pp. 5–6

4.2.3 Pre- and Post-Acquisition Analysis of Disclosure Intimacy

• Intimacy changes are measured as post-minus-pre percentage-point differences and compared with control; multinomial models test condition-by-period differences with user-clustered errors and FDR correction.The control comparison is interpreted descriptively rather than as a causal treatment effect; low intimacy is the regression reference category.Source: §4.2.3 · p. 6

5.1 RQ1: Language Shifts

• Treated posts showed a 42.75% increase in future-focused language and a 10.52% increase in first-person singular pronouns, while the latter declined 7.39% in control posts.Past-focused language declined 5.25% among treated posts, while control posts showed the opposite direction of change.Source: §5.1 · p. 6
• Overall affect, anxiety, sadness, and social language declined 16.65%, 24.43%, 18.26%, and 11.79%, respectively, among treated posts.The corresponding changes in overall affect, anxiety, and sadness were not significant in control posts; discrepancy-related language rose 30.32% in treated posts versus 6.66% in controls.Source: §5.1 · p. 6

5.2 RQ2: Intimacy Shifts

• Low-intimacy disclosures rose for BPD (+13.55 percentage points), depression (+5.14), and anxiety (+3.36), mainly alongside declines in moderate-intimacy posts.High-intimacy disclosures changed comparatively little for these conditions; BD showed no significant change.Source: §5.2 · p. 6, Table 2 · p. 7
• SI-related posts showed a 2.03-percentage-point increase in high-intimacy disclosures, while the control corpus remained relatively stable.Despite this increase, SI-related disclosures did not differ significantly from control.Source: §5.2 · pp. 6–7
• BPD was the only condition with a significant relative decrease in Moderate versus Low intimacy (RRR = 0.553, pFDR < .001).Anxiety, depression, BD, and SI-related disclosures showed no significant differences relative to control.Source: §5.2 · p. 7, Table 3 · p. 7

Social Penetration Theory and Platform Governance

• The findings support considering audience expectations, community norms, affordances, and moderation when interpreting disclosure depth during platform transitions.The authors frame disclosure through perceived rewards and costs of revealing personal information.Source: §6 · p. 7
• Only the relative decline in moderate- versus low-intimacy BPD disclosures was significant; the aggregate analysis cannot establish whether individuals withdrew from intimate exchanges or faced higher disclosure costs.The authors also state that they cannot establish whether users self-censored or isolate the effects of specific governance decisions.Source: §6 · pp. 7–8

Design Implications

• Platforms should accommodate different disclosure preferences, including selective sharing of moderately intimate accounts of coping, relationships, everyday functioning, and treatment.Suggested options include granular audience controls, pseudonymous participation, dedicated mental-health spaces, and consistently enforced guidelines and moderation.Source: §6 · p. 8

Methodological Implications

• The computational framework distinguishes low, moderate, and high intimacy and combines that measure with LIWC analysis to compare disclosure depth across conditions and periods.First-person language can increase while disclosure shifts toward lower intimacy, illustrating that linguistic characteristics and disclosure depth can change differently.Source: §6 · p. 8
• A control corpus provides a baseline for assessing whether mental-health disclosure changes differ from broader communication patterns on the platform.The framework may inform future computational investigations of how platform-level changes relate to stigmatized experiences across online environments.Source: §6 · p. 8

• Disclosure depth does not measure reciprocal relationship development or perceived rewards, costs, responsiveness, or interpersonal closeness, so aggregate shifts cannot establish individual-level social depenetration.Source: Limitations · p. 8
• Aggregate pre/post patterns cannot distinguish changed disclosure by continuing users from changes in who remained active, reduced activity, or left Twitter.The observed changes may reflect changes in disclosure behavior, user composition, or both.Source: Limitations · p. 8
• The dataset includes public Twitter disclosers, while post-acquisition disclosure choices and motives such as trust, impression management, or social validation may shape the observed content and intimacy.Source: Limitations · pp. 8–9
• API-accessible posts are not an exhaustive record; access restrictions challenge reproduction, and deleted posts or suspended accounts may alter historical representation.Source: Limitations · p. 9
• Keyword retrieval phrases can contribute directly to LIWC scores, while changing phrase prevalence and figurative or unverified expressions may affect the retrieved linguistic patterns.Source: Limitations · p. 9
• General Twitter posts are not necessarily comparable to personal disclosures, and their concentration of low-intimacy posts limits their usefulness as a baseline for disclosure depth.Source: Limitations · p. 9
• The post-acquisition period is shorter than the pre-period, so temporal trends, seasonality, contemporaneous events, and evolving disclosure practices may contribute to observed differences.The study does not establish whether the patterns persist, intensify, or diminish over time.Source: Limitations · p. 9

• The analysis used previously collected public Twitter posts without direct interaction, and no institutional ethics review approval or formal exemption determination was obtained.Source: Ethical Considerations · p. 9
• Users may not anticipate that publicly accessible personal experiences will be analyzed at scale, raising privacy and interpretation concerns.Source: Ethical Considerations · p. 9
• Disclosure intimacy is not a measure of mental health severity or clinical status, and the findings should not support conclusions about individuals or harmful surveillance and profiling.Source: Ethical Considerations · p. 9

• AI-assisted language tools, including Grammarly, ChatGPT, and Claude, were used for grammar and readability editing.Source: AI Involvement Disclosure · p. 9
• The disclosed AI use was limited to language editing rather than study design or analysis.Source: AI Involvement Disclosure · p. 9
• The authors conducted the study design, analyses, interpretations, and experiments.Source: AI Involvement Disclosure · p. 9

A Appendix

• Table A1 lists search phrases and collection methods for five mental health samples and the control sample, with disclosure searches spanning January 2018 to February 2023.Source: A Appendix · p. 13, Table A1 · p. 14
• Table A2 reports treated and control pre/post LIWC comparisons, including mean scores, relative changes, net changes, and between-group tests.Source: A Appendix · p. 13, Table A2 · p. 14
• Table A3 provides the codebook for disclosure-intimacy classification.It includes definitions, criteria, examples, and a labeling decision rule.Source: A Appendix · p. 13, Table A3 · p. 15
• The appendix marks Bonferroni-adjusted significance thresholds at p < .05, p < .01, and p < .001.Source: A Appendix · p. 14