We asked a question nobody else was asking:
How do AI personalities actually work? Not how to prompt them. Not how to fine-tune them. How does personality emerge in systems trained on human language?
We built Bladerunner - an experimental platform named after the film's Voight-Kampff test for distinguishing human from machine.
Across 3,352 test cases and 75,709 completions, we tested four frontier models: Claude, GPT-4, DeepSeek, and Gemini. Different companies. Different training data. Different architectures. They converged on the same personality structure (r > 0.90; BFI r = 0.979).
The first cross-provider study of AI personality at scale.
Chomi is built on what we found.
Peer-reviewed research on SSRN
We present a methodology for measuring personality stability in large language models at scale. The Bladerunner platform assigns OCEAN personality profiles to AI systems, then evaluates them using validated psychometric instruments - the same questionnaires used in clinical and research psychology.
Read on SSRN →We propose that personality structure functions as a latent attractor in the space of possible representations learned from language - a stable configuration that training trajectories reliably reach regardless of architecture or provider. Personality geometry is implicit in human language itself.
Read on SSRN →AI personalities drift. Rachel prevents it. A two-speed architecture based on dual-process cognitive theory - fast deterministic correction, slow deliberative analysis. Rachel is Cloudflare for conversational AI.
Read White Paper →