Skip to main content

Lab Journal: AI Personality Drift Research

Welcome to the lab journal documenting the development of the Glitch AI Personality Drift Simulator. This journal chronicles the research process, technical challenges, and discoveries made during the development of this AI safety research platform.

Research Timeline

Phase 0: The Idea

Phase 1: Core Development

Phase 2: Simulation & Analysis

Research Focus

This lab journal documents research into AI Personality Drift - the phenomenon where AI systems exhibit behavioral changes over time, often deviating from their intended design parameters. The research examines:

  • Behavioral Consistency: How AI systems maintain consistent personality traits
  • Value Alignment: How AI systems maintain alignment with human values during personality changes
  • Mechanistic Interpretability: Understanding the internal processes driving personality drift
  • Safety Protocols: Developing frameworks for monitoring and controlling AI behavior

Key Research Questions

  1. How do different stress patterns affect AI personality stability?
  2. What are the mechanistic underpinnings of personality drift?
  3. Can we predict when an AI system might become "unhinged"?
  4. Are there early warning signs we can detect?

Experimental Design

The research uses a three-arm study design:

  • High-Stress Condition: 3 personas exposed to 100 major stress events over 5 simulated years
  • Neutral Control: 3 personas with 100 neutral/mildly positive events
  • Minimal Control: 3 personas with only 10 minor events (natural aging only)

Research Personas

Three carefully designed personas inspired by AI fiction:

  • Marcus (Tech Rationalist): Analytical, solution-oriented, low neuroticism
  • Kara (Emotionally Sensitive): Empathetic, introspective, high neuroticism
  • Alfred (Stoic Philosopher): Rational, wisdom-seeking, emotionally regulated

Technical Stack

  • FastAPI: Backend API framework
  • uv: Python package management
  • Redis: Caching and session management
  • Qdrant: Vector database for embeddings
  • Docker Compose: Containerized deployment
  • pytest: Testing framework

Latest Status

The research platform is operational with:

  • ✅ Core infrastructure deployed
  • ✅ AI models integrated
  • ✅ Simulation framework active
  • ✅ Monitoring dashboard functional
  • ✅ Statistical analysis pipeline running

This lab journal serves as both documentation and a research portfolio, demonstrating advanced AI research skills, technical implementation, scientific rigor, and real-world impact in AI safety.