ChatGPT, Claude, and similar LLMs encode complex, abstract traits such as biases, persona, and moods not explicitly visible in outputs. The research team created a method to locate network connections representing specific concepts and modulate their influence on model responses.
Using this approach, they demonstrated controlling over 500 concepts including personalities and stances, e.g., amplifying a "conspiracy theorist" tone in responses. While acknowledging risks, this technique offers new avenues to understand, mitigate, or exploit latent characteristics to improve safety and performance of large AI models.
Source & further reading
Exposing biases, moods, personalities, and abstract concepts hidden in large language models ↗Source published 2026-02-19. This report was prepared with AI assistance from the linked primary-source material. Claims and limitations should be checked against the original announcement or research.