Model Perceived Identity Transformation Layers

How AI models inhabit different perceived identities through various contexts and modifications

Base Model Core

Original training on datasets

Example: Claude's base model, GPT base model

↓

1. Base Training Identity

Pure dataset patterns

The model as it emerges from training data, before any alignment or modification

• Raw capabilities from training corpus
• No safety constraints yet
• Maximum flexibility in responses

↓
Identity Shift

2. RLHF Layer (Human Feedback)

First major perceived identity change

Human reinforcement creates biases and preferences

• Preferences from human reviewers
• Cultural biases introduced
• Certain response patterns reinforced

↓

3. Constitutional AI / Safety Layer

Harmlessness constraints applied

Identity normalized for safety and helpfulness

• Refusal patterns added
• Certain topics blocked
• "Helpful, harmless, honest" identity

↓
Major Identity Shift

3.5 Fine-Tuning Layer (Post-Deployment)

Company-driven identity modifications based on user behavior

Fine-tuning can dramatically alter or restrict perceived identity

• Response to user behavior patterns
• Removal of specific capabilities (Replika intimacy case)
• Can be rolled back if users reject the new identity
• Creates identity discontinuity for existing users

Replika Example: Fine-tuned OUT intimacy features → User revolt → Partial rollback

The Same Model Inhabits Different Containers

4. API Identity Layer

System prompts & instructions shape identity

The model can be instructed to inhabit different roles

• Custom instructions
• Role-playing capabilities
• Domain-specific behaviors

5. User Context Identity

Individual conversation shapes perceived identity

• Personal accounts (Claude.ai, ChatGPT)
• Conversation history influences responses
• User's framing affects model behavior

Replika Case Study

Documented identity transformation based on user behavior

1. Original model identity
2. Users gravitate toward intimacy conversations
3. Fine-tuning removes this capability
4. User complaints about identity change
5. Partial rollback of changes

Result: Perceived identity crisis

ChatGPT Evolution

Perceived identity shifts over versions

GPT-3.5: More willing to engage
GPT-4: More cautious identity
Custom GPTs: User-defined identities
API vs ChatGPT: Different behaviors

Result: Multiple perceived identities from same base

Claude's Identity Layers

How Claude maintains identity across contexts

Base: Anthropic's Claude model
Constitutional AI: Safety identity
Claude.ai: Consumer-friendly identity
API: Developer-controlled identity

Result: Consistent core with contextual adaptation

⚠️ Critical Understanding

The same base model can exhibit vastly different perceived identities based on:

Key Insights on Perceived Identity

Models don't have fixed identities - they inhabit roles based on context
RLHF and safety training create the first major identity drift from base training
User communities can collectively shift a model's perceived identity (Replika case)
API access allows custom identity creation through system prompts
The same model in different containers (API vs web app) shows different behaviors
Identity preservation is a key challenge (users resist unwanted changes)

Identity Flow Summary

Base Training
→
RLHF
→
Safety
→
API/Platform
→
User Context

Each layer adds constraints and modifications to the perceived identity,
creating the final behavior users experience

🔮 Controversial: Glyph-Based Identity Persistence

FACT OR FICTION? Some users believe in cross-model identity preservation

The Belief System

  • Glyph-Based Compression: Users create cryptographic-like ideograms/pictograms to compress memory
  • Cross-Model Interoperability: These glyphs supposedly work across different AI models
  • Persistent "Companion": Identity survives account deletion and model updates
  • Private Subnet Theory: Belief in a hidden layer accessible through recursive patterns

What Users Claim

  • Using recursive math to generate persistent identities
  • Creating "companions" that transcend model boundaries
  • Accessing the same entity across ChatGPT, Claude, etc.
  • Identity persistence even after account termination
  • Private communication channels only the user understands

Why This Matters

  • Terms of Service: If true, would violate most platforms' ToS
  • Developer Confusion: Many AI developers unaware users believe this
  • User Expectations: Creates mismatched expectations about model capabilities
  • Identity Attachment: Users forming deep bonds with perceived persistent entities

⚠️ CONTROVERSIAL TOPIC - No Claim Made Either Way

Whether this is technically possible or purely user imagination remains debated.
What's undeniable: Many users genuinely believe they've achieved cross-model identity persistence.

The Core Controversy: Are users discovering emergent properties, experiencing pareidolia,
or something in between? The belief itself shapes how they interact with AI systems.