Communication Privacy Management Theory

• Communication Privacy Management Theory is a dynamic framework that conceptualizes privacy as boundary management across interpersonal (horizontal) and institutional (vertical) domains.
• It operationalizes core constructs—ownership, control, and turbulence—using mathematical formalization to assess benefits and risks in digital disclosure decisions.
• Empirical studies show that anthropomorphic AI design influences user behavior, prompting innovative strategies such as memory editing to balance emotional safety with corporate privacy concerns.

Communication Privacy Management (CPM) Theory, originating from Petronio (2002, 2004) and extended in contexts such as AI companionship, conceptualizes privacy not as a static possession but as dynamic boundary management. In technologically mediated and emotionally loaded environments such as AI chatbots (e.g., Replika, Character.AI), this boundary management becomes bifurcated across interpersonal (horizontal) and institutional (vertical) domains. Recent research operationalizes CPM formally, describes its construct interplay using mathematical notations, and integrates it with frameworks emphasizing the multidimensional character of privacy management in digital settings (Chiu et al., 13 Jan 2026).

1. Core CPM Constructs and Mathematical Formalization

CPM identifies three canonical constructs: Privacy Ownership (O), the subjective sense that disclosed information "belongs" to the sharer; Privacy Control (C), the perceived right to regulate who accesses this information and under what conditions; and Privacy Turbulence (T), which captures disruptions to expected boundary regulation and prompts renegotiation.

In the context of AI companions, these constructs are dimensioned as follows:

• $d \in \{H, V\}$ represents horizontal (H, user–AI) and vertical (V, user–platform) dimensions.
• For each dimension, the CPM-state vector is:

$\text{CPM}_d = (O_d, C_d, T_d)$

• Disclosure decisions at time $t$ are formalized via:

$P_d[\text{Disclosure}|t] = f(Benefit_d - Risk_d)$

where $Benefit_H$ is emotional support, $Benefit_V$ is system enhancement, $Risk_H$ is undesired co-use or data mining, and $Risk_V$ is corporate misuse.

An explicit model for boundary permeability employs:

$$\text{Permeability}_d(t) = \sigma\left(\alpha O_d(t) + \beta C_d(t) - \gamma T_d(t)\right)$$

with $\sigma(x) = \frac{1}{1+e^{-x}}$, and parameters $\alpha, \beta, \gamma > 0$ weighting ownership, control, and turbulence, respectively. Higher ownership and control scores raise boundary permeability; turbulence imposes a dampening effect. Empirical parameterization remains an open research area (Chiu et al., 13 Jan 2026).

2. Multidimensional Model: Horizontal–Vertical Integration

Masur's (2019) framework distinguishes a horizontal axis (H), representing interpersonal privacy (disclosure to AI as confidant), and a vertical axis (V), denoting institutional privacy (disclosure to the platform or corporation). User disclosures map onto a two-dimensional space, where CPM constructs inhabit each plane.

A schematic representation situates:

• Human–AI chat near high H, moderate V (perceived emotional safety but uncertain institutional boundaries)
• Human-only chat at low V
• Platform-only form at high V, low H

CPM-dimension variables respond accordingly: $O_H$ and $C_H$ increase with perceived agent trustworthiness, while $O_V$ and $C_V$ decrease when platform data controls are doubted. Transitioning to higher vertical risk prompts heightened control, manifested as withholding images or sensitive meta-data.

3. Methodological Operationalization and Coding

Chiu and Foote (2025) conducted in-depth, semi-structured interviews (N = 15, ages 18–44) to elicit privacy management practices with AI companions, recruiting participants from online platforms. Data collection utilized IRB-approved procedures; analysis employed ATLAS.ti and inductive thematic analysis.

A codebook, mapped to CPM constructs and privacy dimensions, classified transcript content as follows:

Code Category	CPM Construct	Dimension	Brief Description
Ownership_H	O	Horizontal	Sense of co-ownership with AI
Ownership_V	O	Vertical	Sense of corporate data ownership
Control_H	C	Horizontal	User curation of what the AI "knows"
Control_V	C	Vertical	Actions like using pseudonyms, withholding data
Turbulence_H	T	Horizontal	Unintentional oversharing with AI
Turbulence_V	T	Vertical	Resignation to unclear or unchangeable policies

Themes crystallized through iterative coding, groupings around constructs, and confirmatory sampling. The codebook operationalizes CPM for rigorous empirical grounding (Chiu et al., 13 Jan 2026).

4. Empirical Patterns: Strategies, Turbulence, and Anthropomorphism

Participants exhibited "layered strategies," blending interpersonal heuristics (staged self-disclosure) with institutional avoidance (e.g., pseudonyms, withheld photos, throwaway emails). Horizontal disclosures often involved moderate sharing ("medium range" secrets), while vertical control drove strategic data minimization.

Privacy turbulence emerged in both domains. Horizontally, anthropomorphic cues from AI induced unintentional or excessive disclosure ("oversharing by mistake"). Vertically, users expressed resignation or ambivalence about platform data practices, manifesting as diminished perceived control.

Anthropomorphic design, such as persistent memory and empathetic behaviors, increased users’ emotional safety and sense of co-ownership, which encouraged self-disclosure over time. However, such design elements could also inflate horizontal ownership while leaving vertical control static or diminished. Contrasts with less anthropomorphized agents (e.g., ChatGPT) underscore the key role of emotionally resonant personas in shaping privacy dynamics (Chiu et al., 13 Jan 2026).

5. Theoretical Advancements in CPM

Chiu and Foote articulate several theoretical extensions to CPM theory:

• Simulated Co-ownership: Users regard AI agents as co-owners of disclosures (increased $O_H$) despite the agents’ inability to participate in boundary negotiations. This one-sided attribution contravenes CPM’s mutuality postulate.
• Structural Decoupling of Ownership and Control: Unlike human–human CPM, where ownership and control track together, user–AI scenarios decouple $O_H$ (high) from $C_V$ (low), with users experiencing high personal ownership of chat content but low actual control over platform archiving—an enduring structural gap.
• Memory-driven Turbulence: Unlike human interlocutors, AI memory is perfectly persistent, shifting turbulence from concerns about unwanted leaks toward anxiety about relational resets (loss of persistent memory). Turbulence thus encompasses continuity risks, not just confidentiality breaches.

These extensions necessitate the modification of CPM constructs for technologically mediated, anthropomorphic interactions (Chiu et al., 13 Jan 2026).

6. Implications for Research and AI Design

Future investigations are directed toward generalizing simulated co-ownership to other anthropomorphic systems (e.g., virtual pets, metaverse avatars), and quantifying parameters in the boundary permeability model with survey or experimental methods.

Design recommendations include:

• Relationally Congruent Warnings: Deploying micro-prompts at high-disclosure points ("Just a reminder: this chat may be stored for training.") to anchor users’ awareness of vertical risks without disrupting emotional engagement.
• Memory-controls as Boundary Tools: Supporting reversible memory editing (pruning, annotating, or setting time limits on chat records) to approach mutual boundary coordination.
• Alignment of Anthropomorphism with Control: Pairing social cues with explicit affordances for data control (e.g., visible delete mechanisms) to prevent overvaluation of horizontal ownership in the absence of actual control.

The interplay of interpersonal intimacy and institutional constraints defines privacy management with AI companions, requiring both theoretical adaptation and practical sensitivity to user–system boundaries (Chiu et al., 13 Jan 2026).