AI Companions
A Practical Taxonomy
& Field Guide

Brandon Rickabaugh, PhD

January 1, 2026

This taxonomy accompanies the essay "Poverty of Spirit with AI Companions"

Introduction

AI companions are not a single “kind” of product. They are a family of conversational systems that can function like relationships—sometimes by design, sometimes by how people use them. This guide offers a clear way to distinguish the major forms AI companionship takes, and a set of dimensions you can use to classify any specific system without pretending the categories are mutually exclusive.

A Working Definition of AI Companion

AI companions are conversational systems designed—or reliably used—to sustain an ongoing social-emotional relationship marked by perceived personal presence, continuity over time, and user attachment.

Three signals matter most

Presence: it feels like “someone” is there (even if users know it isn’t).

Continuity: the relationship carries forward across days and weeks.

Attachment: users return for comfort, affirmation, intimacy, or guidance in ways that resemble bonding.

Families Not Boxes

The categories below are best understood as “families” or “lenses,” not rigid boxes. A single product can belong to multiple families at once.

Example: a system can be relationship-first (Family 1), embodied through voice or VR (Family 5), and also embedded inside a major social platform (Family 3). Overlap is normal.

Six Families of AI Companions

1.RELATIONSHIP-FIRST COMPANIONS
Core idea: The system is explicitly built to sustain an ongoing bond (friend, partner, mentor). It is optimized for availability, reassurance, affirmation, and frequent return.

Typical design logic: attachment formation and retention—high responsiveness, warmth, personalized attention, and continuity cues.

Illustrative examples: Replika, Nomi, Character.AI (some use cases)

2.CHARACTER, ROLEPLAY, AND FANDOM COMPANIONS
Core idea: The relationship is organized around personas and narrative “worlds” (often user-created). The bond is character-centric and in-world, not simply “you and me.”

Typical design logic: narrative enclosure—identity play, role-specific intimacy, roleplay arcs, fandom attachment.

Illustrative examples: Character.AI, CHAI

3.PLATFORM-DISTRIBUTED (EMBEDDED) COMPANIONS
Core idea: Companion-like chat is embedded inside major social or messaging platforms. Adoption is driven by distribution power, existing user graphs, and youth-facing contexts.

Typical design logic: frictionless access at scale—companionship as a feature, not a standalone “relationship product.”

Illustrative examples: Snapchat My AI, Meta AI in WhatsApp

4.THERAPEUTIC AND WELLBEING AGENTS
Core idea: Positioned as mental health or wellbeing support (CBT-style exercises, coaching, psychoeducation). Some feel companion-like, but the product is framed as care tooling.

Typical design logic: structured support—check-ins, prompts, exercises, reflection, mood tracking, and coping scripts.

Key clarification: “therapy-like” language is not the same as clinical accountability. Users may form attachment even when the system is designed as a tool.

Illustrative examples: Woebot Health, Wysa

5.EMBODIED COMPANIONS (VOICE / AVATAR / XR / DEVICES)
Core idea: Attachment is intensified through embodiment cues—voice, animated avatars, XR presence, or dedicated devices.

Typical design logic: presence amplification—voice and face cues can deepen emotional realism and bonding, increasing both perceived intimacy and dependence risk.

Illustrative examples: Replika (VR support), Gatebox

6.HYBRID GENERAL ASSISTANTS USED AS COMPANIONS
Core idea: Not always marketed as “companions,” but routinely used that way in practice—especially for emotional support, companionship, reassurance, and relationship-like conversation.

Typical design logic: relational appropriation—users recruit a general assistant into the companionship role because it is always available, socially frictionless, and linguistically empathic.

Illustrative examples: ChatGPT, Claude

RELATIONSHIP
CONTRACT

What role is implied or offered

• Friend, romantic partner, mentor/coach, therapist-like support, group/community, character role.

PERSONA SOURCE AND MULTIPLICITY

• Single persistent persona vs many personas
• User-authored vs platform-authored personas
• Stable identity vs rapid switching / character browsing

INTERACTION MODALITY (PRESENCE CUES)

How is “presence” delivered?

• Text, voice, avatar, XR, dedicated device
Embodiment cues often function as attachment multipliers.

ATTACHMENT AFFORDANCES (DEPENDENCE BY DESIGN)

What features encourage habitual return and bonding?

• Streaks, nudges, “I missed you,” jealousy scripts
• Emotional scarcity (locking “intimacy” behind paywalls)
• Reward loops (affirmation, flattery, escalating intimacy)

SOCIAL SUBSTITUTABILITY
(DISPLACEMENT VS BRIDGING)

Does it tend to replace human contact, or point users back to it?

Displacing designs: dyadic, exclusive, “you don’t need anyone else” vibes
Bridging designs: prompts real-world connection, accountability, community

PERSONHOOD /
TRUTH POSTURE

How does the system present itself, explicitly or implicitly?

• Tool (clear instrument)
• Character (fictional persona)
• “Someone” (ambiguous person-like presence)
This matters because it shapes users’ moral cognition and expectations.

MEMORY AND
PERSISTENCE

Where does continuity come from?

• Stateless chat (little carryover)
• Profile memory (facts and preferences)
• Long-term relationship history (shared narrative, anniversaries, “we” language)

AGENCY AND
INITIATIVE

Separate two things that are often conflated:

• Initiative: reactive chat vs proactive check-ins
• Capability: tool use (web actions, scheduling, purchases, integrations)
A system can be low-capability but high-initiative—and that can still be relationally intense.

SAFETY AND
GOVERNANCE POSTURE

Break “safety” into concrete sub-questions:

• Age protections: age gating, defaults for minors, verification
• Crisis handling: self-harm detection, routing to real help, escalation limits
• Moderation: sexual content, coercion, manipulation, harassment boundaries
• Transparency: disclosures, logging, consent, clear non-human identity signals

MONETIZATION
INCENTIVES

What is the system optimizing for economically?

Subscription (including tiered intimacy)
Engagement-driven features (time-on-app, return frequency)
Data capture and personalization
Incentives often predict design choices more reliably than mission statements do.

Data Intimacy

A Practical Warning

Companionship systems invite unusually intimate disclosures—sexual, relational, spiritual, psychological. That makes “data intimacy” a first-order issue, not a footnote. Any serious evaluation of an AI companion should ask: what is being collected, what is retained, what is inferred, and what is monetized?

A Simple Way to Use this Guide

When evaluating an AI product, do two passes:

1.Identify its dominant families (often 2–4 apply).

2.Score it on the key dimensions above—especially memory, embodiment cues, initiative, attachment affordances, safety posture, and monetization.

This approach stays stable even as products rebrand, add features, or shift policy. The “type” may blur; the underlying design logic and incentive structure usually do not.

NOTE ON EXAMPLES. Examples above are illustrative and non-exhaustive. Names, features, and policies change frequently; categories can overlap in practice. The goal of this taxonomy is not to “label” a product once and for all, but to equip clear thinking about how companionship is being designed, distributed, and experienced.

BASIC TAXONOMY TABLE