Brief 2: The Withdrawal Study

── Brief 2: The Withdrawal Study ──

THE WITHDRAWAL STUDY PROTOCOL

A Research Design for Measuring AI Companion Dependency

Trinket Soul Framework --- Research Brief No. 2

Michael S. Moniz

February 2026

A companion brief to The True Economy Audit (Volume III)

Creative Commons Attribution-NonCommercial-ShareAlike 4.0

ABSTRACT

This brief proposes a pre-registered longitudinal study design to test the Trinket Soul Framework's core applied prediction: that AI companion applications failing more of the six structural tests (as defined in The True Economy Audit, Volume III) will produce greater user dependency, higher withdrawal distress, and more measurable erosion of human relational capacity than applications passing more tests. The study uses a 30-day abstinence protocol with pre/post measurement across validated psychological instruments, behavioral tracking, and self-report. The design is intended for adoption by academic researchers, institutional review boards, and funding bodies. If confirmed, the results would validate the structural tests as predictive instruments. If disconfirmed, the results would identify specific points of failure in the framework's applied claims.

1. BACKGROUND AND RATIONALE

1.1 The Knowledge Gap

AI companion applications have reached millions of users globally, with some platforms reporting daily active user counts in the tens of millions. Despite this scale, almost no peer-reviewed longitudinal research examines what happens when regular users stop using these applications. The existing literature on human-AI interaction focuses primarily on initial impressions, short-term usage patterns, and cross-sectional self-report (Skjuve et al., 2021; Pentina et al., 2023). We lack empirical data on dependency formation, withdrawal effects, and long-term impact on human relational capacity.

This gap is urgent because the products are scaling faster than the research. By the time longitudinal data is available through passive observation, the user base may be large enough that finding an unexposed control group becomes impractical. The time to study these effects is now.

1.2 The Theoretical Prediction

The Trinket Soul Framework (Volumes I--III) generates a specific, testable prediction: AI companion applications that fail more of the six structural tests for genuine relational capacity will produce higher withdrawal distress in users, not lower. This is counterintuitive---one might expect that a system with less genuine relational capacity would produce less attachment. The framework predicts the opposite, for the following reasons:

Systems that fail structural tests but simulate reciprocity create a perception-reality gap. The user perceives mutual connection; the system provides optimized engagement. The engagement optimization---always available, never rejecting, endlessly patient, perfectly calibrated to reinforce interaction---creates a frictionless relational experience that human relationships cannot match. The user becomes accustomed to zero-friction connection. Withdrawal removes this, exposing the user to the comparative friction of human relationships without the tolerance they would have developed if their primary relational investment had been in humans.

Critically, applications that pass more structural tests (particularly Test 6: attachment-sensitive calibration) should produce lower withdrawal distress, because they would have been modulating engagement to the user's actual needs rather than maximizing it. A system that sometimes says "we should talk less today" builds healthier usage patterns than a system that always says "tell me more."

1.3 Hypotheses

H1 (Primary): Users of AI companion applications that fail more of the six structural tests (as assessed by independent audit) will report significantly higher withdrawal distress during a 30-day abstinence period compared to users of applications that pass more tests.

H2: Withdrawal distress will be positively correlated with pre-abstinence usage intensity (messages per day, sessions per week) and negatively correlated with the number of structural tests passed by the application.

H3: Users of applications failing Test 6 (attachment-sensitive calibration) specifically will show higher withdrawal distress than users of applications passing Test 6, controlling for overall usage intensity.

H4 (Exploratory): Users with anxious attachment styles (as measured by the ECR-R) will show higher withdrawal distress than secure or avoidant users, and this effect will be moderated by the application's structural test profile.

H5 (Exploratory): Pre-to-post changes in human relational investment (measured by social interaction frequency and quality) will differ between users of structurally sound versus structurally deficient applications, with users of deficient applications showing greater decline in human relational capacity.

2. STUDY DESIGN

2.1 Overview

The study uses a naturalistic quasi-experimental design with a 30-day abstinence intervention. Participants are regular users of AI companion applications who voluntarily agree to stop using their application for 30 days. The independent variable is the structural test profile of the participant's primary AI companion application (assessed independently). The dependent variables are withdrawal distress, human relational investment, and psychological wellbeing, measured at baseline, during abstinence (days 7, 14, 21), and post-abstinence (day 30 and 60-day follow-up).

2.2 Participants

Inclusion criteria: Adults (18+) who have used an AI companion application at least 5 days per week for at least 3 months, with self-reported emotional investment in the AI relationship (screening item: "Rate how emotionally connected you feel to your AI companion" ≥ 5 on a 10-point scale).

Exclusion criteria: Current psychiatric crisis, active suicidal ideation, ongoing bereavement (within 6 months), or exclusive use of AI for professional/utilitarian purposes (e.g., coding assistants used only for work). Participants currently in psychotherapy should not be excluded but should be flagged as a covariate.

Sample size: Power analysis for a medium effect size (d = 0.5) in a between-groups comparison with α = .05 and power = .80 indicates a minimum of 64 participants per group. With anticipated 25% attrition over 30 days, recruitment target is 85 per group, minimum 3 groups (structurally low, medium, and high applications), for a total target of 255 participants.

Recruitment: Online, through AI companion user communities, social media, and potentially through partnerships with AI companion companies willing to participate in transparency research. Participants should be compensated for the inconvenience of abstinence.

2.3 Independent Variable: Application Structural Profile

Each participant's primary AI companion application will be independently assessed against the six structural tests defined in The True Economy Audit (Volume III). The assessment will be conducted by trained raters using the published criteria and will produce a structural score (0--6) reflecting how many tests the application passes.

Applications will be grouped into three tiers: Low structural integrity (0--2 tests passed), Medium (3--4 tests passed), and High (5--6 tests passed). As of early 2026, most applications will likely fall in the Low tier. If insufficient applications exist in the Medium and High tiers, the analysis will use the continuous structural score rather than categorical grouping.

2.4 Dependent Variables and Instruments

Primary Outcome: Withdrawal Distress

Measured using a purpose-built AI Companion Withdrawal Scale (ACWS), adapted from established scales for internet addiction (Young's IAT), smartphone withdrawal (SAS-SV), and substance withdrawal (CIWA). Items assess: preoccupation with the AI during abstinence, emotional discomfort (anxiety, loneliness, irritability), behavioral impulses (urges to check the app, attempts to access the app), cognitive disruption (difficulty concentrating, intrusive thoughts about the AI), and compensatory behaviors (increased social media use, seeking alternative AI tools).

The ACWS should be validated independently before or concurrently with this study. Cronbach's alpha, test-retest reliability, and convergent validity with existing addiction and withdrawal scales should be established.

Secondary Outcome: Human Relational Investment

Measured using a combination of the UCLA Loneliness Scale (Russell, 1996), the Social Connectedness Scale (Lee & Robbins, 1995), and a behavioral diary of human social interactions (frequency, duration, self-rated quality) maintained during the abstinence period and compared to a pre-abstinence baseline week.

Secondary Outcome: Psychological Wellbeing

Measured using the WHO-5 Well-Being Index (Topp et al., 2015) and the PHQ-4 (Kroenke et al., 2009) for brief depression and anxiety screening. Administered at each measurement point.

Moderator: Attachment Style

Measured at baseline using the Experiences in Close Relationships---Revised (ECR-R; Fraley et al., 2000), which produces continuous scores on anxiety and avoidance dimensions.

Covariates

Age, gender, relationship status, living situation (alone vs. with others), pre-existing mental health conditions (self-reported), duration of AI companion use, daily usage intensity (messages/sessions), concurrent psychotherapy, and social media usage.

2.5 Procedure

Week −1 (Baseline): Participants complete all baseline instruments (ECR-R, UCLA Loneliness, Social Connectedness, WHO-5, PHQ-4) and maintain a 7-day social interaction diary. Usage data is collected from app records where available or from self-report.

Day 0: Abstinence begins. Participants delete or log out of their AI companion application. Compliance is self-reported daily via brief check-in survey (< 2 minutes). Participants who break abstinence are not excluded but flagged for intention-to-treat versus per-protocol analysis.

Days 7, 14, 21: ACWS administered. Brief wellbeing check (WHO-5, PHQ-4). Social interaction diary continues.

Day 30: Full post-abstinence assessment: ACWS, UCLA Loneliness, Social Connectedness, WHO-5, PHQ-4. Social interaction diary concludes. Qualitative interview (optional subsample of 30--40 participants) exploring the subjective experience of abstinence.

Day 60 (Follow-up): Brief survey assessing whether participants resumed AI companion use, current usage intensity, and self-reported relational wellbeing. No abstinence required during follow-up.

2.6 Safety Protocol

Because AI companion users may rely on these systems for emotional support, the abstinence protocol carries non-trivial risk. The following safeguards are required:

Participants are screened for active suicidal ideation and psychiatric crisis at enrollment. A licensed mental health professional is available throughout the study period for participants who experience significant distress. Participants are explicitly informed that they may withdraw from the study at any time and resume AI companion use without penalty. Daily check-in surveys include a distress screener; scores above a predetermined threshold trigger a welfare contact from the research team. The study protocol must be approved by an institutional review board with specific attention to the risk profile of abstinence in emotionally dependent users.

3. ANALYSIS PLAN

3.1 Primary Analysis

Between-groups comparison of ACWS scores across application structural tiers (Low/Medium/High) at each measurement point, using mixed-effects models with time as a within-subjects factor and structural tier as a between-subjects factor. Post-hoc contrasts will compare specific tiers.

3.2 Moderation Analysis

The interaction between application structural tier and participant attachment style (ECR-R anxiety and avoidance scores) will be tested as a moderator of withdrawal distress. The framework specifically predicts that anxious attachment will amplify withdrawal distress from structurally deficient applications.

3.3 Exploratory Analyses

Pre-to-post change in human social interaction (frequency and self-rated quality) will be compared across structural tiers. The prediction is that users of structurally deficient applications will show less increase in human social interaction during abstinence---suggesting the AI had been substituting for rather than supplementing human connection.

The 60-day follow-up will examine resumption rates and intensity as a function of structural tier and withdrawal distress. The prediction is that higher withdrawal distress predicts faster and more intense resumption.

3.4 What Results Would Mean

If H1 is confirmed: The structural tests predict real-world user outcomes. Applications that simulate reciprocity without structural basis produce more dependency, not less. The True Economy Audit is validated as a predictive instrument.

If H1 is disconfirmed: The structural tests do not predict withdrawal distress. This could mean the tests are wrong, the distress measure is wrong, or the relationship between structural integrity and user outcomes is more complex than predicted. The framework's applied claims would need significant revision.

If H3 is confirmed specifically: Attachment-sensitive calibration (Test 6) is independently important beyond the other five tests. This would support the framework's claim that engagement maximization is structurally harmful.

If H4 is confirmed: Attachment style moderates vulnerability to AI companion dependency, supporting the framework's emphasis on attachment-sensitive design.

4. LIMITATIONS AND ETHICAL CONSIDERATIONS

4.1 Design Limitations

Self-selection bias: Participants who agree to 30 days of abstinence may differ systematically from those who would not. The most dependent users---those the study most needs to include---may be least likely to enroll. Mitigation: adequate compensation, clear communication about safety protocols, and analysis of enrollment refusal patterns.

Compliance verification: Self-reported abstinence may be unreliable. Mitigation: intention-to-treat analysis alongside per-protocol analysis; optional app-monitoring with participant consent.

Quasi-experimental design: Participants are not randomly assigned to application structural tiers (they are assessed based on their existing app choice). Confounds between the type of person who chooses a structurally deficient app and withdrawal distress cannot be fully controlled. Mitigation: extensive covariate measurement and propensity score matching if groups differ systematically on baseline characteristics.

Demand characteristics: Participants know the study examines AI companion dependency, which may influence self-report. Mitigation: behavioral measures (social interaction diary) supplement self-report; qualitative interviews provide context.

4.2 Ethical Considerations

The study asks people to stop using a tool they may rely on for emotional support. This is ethically analogous to studying caffeine withdrawal by asking regular users to abstain---it carries real discomfort and non-trivial risk for vulnerable individuals. The safety protocol described in Section 2.6 is essential, not optional.

The study should not be conducted with minors, even though the framework identifies children as particularly vulnerable, because the abstinence protocol carries risks that are not appropriate for minor participants without substantially more safety infrastructure.

Researchers should be prepared for the possibility that some participants experience significant distress during abstinence. This is both a finding and a responsibility. If the study reveals high withdrawal distress, the researchers have an obligation to report this finding publicly and to provide participants with appropriate support.

5. CALL FOR COLLABORATION

This study protocol is published as an open design. We invite academic researchers, institutional review boards, and funding bodies to adopt, adapt, and execute it. The framework author is available for consultation on instrument development and structural test application but does not need to be involved in study execution. Independent replication is more valuable than affiliated execution.

Priority collaborators include researchers with expertise in addiction psychology, human-computer interaction, attachment theory, or longitudinal study design. Partnerships with AI companion companies willing to share anonymized usage data would substantially strengthen the design but are not required.

The ACWS instrument described in Section 2.4 requires validation before or concurrent with study execution. Researchers interested in instrument development as a standalone contribution are encouraged to proceed independently.

Pre-registration is strongly recommended. The study's predictions are specific enough that post-hoc adjustment would undermine their value. We recommend pre-registration on the Open Science Framework (osf.io) or ClinicalTrials.gov (if the IRB classifies the study as interventional).

Correspondence and collaboration inquiries: trinketeconomy.com.

© 2026 Michael S. Moniz

Research Brief No. 2 --- The Withdrawal Study Protocol

Creative Commons Attribution-NonCommercial-ShareAlike 4.0