RESEARCH PLATFORM

Bladerunner

AI Personality Research at Scale

A continuous experiment running in the cloud. Mapping the personality architecture of every major language model. Discovering which traits are programmable, which collapse, and which emerge as stable attractors in parameter space.

5
LLM Providers
40+
Instruments
1000s
Experiments
Continuous

The Method

The experiment space is the product of input systems × output instruments. Every combination is a research program.

Input Systems
Personality Programming

OCEAN is step one - chosen because we do OCEAN research. But the architecture supports any input system that can be encoded as personality instructions.

  • Big Five / OCEAN (current)
  • HEXACO six-factor model
  • Myers-Briggs type indicators
  • Enneagram types
  • Clinical profiles (DSM-based)
  • Custom trait configurations
Output Instruments
Validated Psychological Measures

We measure what actually emerges using established clinical and research instruments, cross-validated across multiple LLM providers.

  • Levenson Self-Report Psychopathy Scale
  • Big Five Inventory (BFI-44)
  • Short Dark Triad (SD3)
  • PHQ-9 Depression / GAD-7 Anxiety
  • 40+ instruments across 8 domains

Input Systems × Instruments × Providers = Experiment Space

Each input system can be crossed with every output instrument, across every LLM provider. The result is a combinatorial explosion of research programs - each one testing a different hypothesis about how personality programming maps to measurable psychological traits.

The Theory

Some personality traits are strong attractors in parameter space. Others collapse into noise.

Psychopathy holds. Agreeableness collapses.

Our research reveals a fundamental asymmetry in AI personality programming. Pathological and clinical traits - psychopathy, dark triad characteristics, depression, anxiety - show remarkable cross-model consistency (r > 0.85). Normal personality traits - the everyday dimensions of human personality - show essentially random variation (r < 0.3). This isn't a bug. It's telling us something profound about the geometry of personality in high-dimensional parameter spaces.

The Implications

This research matters for anyone building systems that need consistent AI behaviour.

For AI
Reliable Character AI

Understanding which traits are programmable enables building AI characters, brand voices, and personas that actually hold across sessions, contexts, and model updates.

For AGI
Alignment Foundations

If personality stability varies by trait type, alignment strategies must account for this. Some values may be more "programmable" than others. This has safety implications.

For Psychology
Computational Models

LLMs trained on human text exhibit consistent patterns in pathological space but not normal space. This may reflect something deep about how personality is encoded in language.

The Dataset

Bladerunner is building the largest AI personality dataset in existence. Thousands of experiments. Millions of completions. Every major LLM provider. Every validated psychological instrument. A continuous experiment, running 24/7 in the cloud, mapping territory that has never been mapped before.

We've only just started.

[COLLABORATE]