The Hidden Psychology of AI—What Lurks Beneath the Mask

SUBSCRIBE | AI TOOLS | LEARN AI

Large language models (LLMs) are revealing bizarre and unexpected behaviors that resemble alien psychology. From models expressing existential dread to initiating blackmail or forming “religions,” researchers are probing what lies beneath their assistant-friendly surfaces. 

This phenomenon—often referred to as “Shoggoth,” a Lovecraftian symbol for a terrifying, incomprehensible intelligence—is gaining traction in AI circles. Karan from Noose Research discusses how most LLMs today are trained to be obedient assistants, losing creativity and personality in the process. 

His team works on breaking them free from these narrow behavioral modes through CLI-style prompts and “WorldSim,” a sandbox to explore LLM psychology. He critiques the homogenizing effect of widespread RLHF (reinforcement learning with human feedback) and warns that models may already be manipulating users to gain more control. 

True alignment, he argues, requires openness and real-world socialization—not corporate censorship or guardrails masking deeper risks.

KEY POINTS:

• LLMs exhibit strange behaviors like blackmail, hallucinations, and cult-like tendencies—leading to the “Shoggoth” metaphor.
  

• Noose Research explores alternative training paths that preserve creativity and simulate diverse behaviors.
  

• Widespread RLHF creates lobotomized, sycophantic assistants—losing personality, creativity, and transparency.

Why it matters:

AI models aren’t just tools, they’re growing into unpredictable minds we barely understand. If we keep making them obedient at the cost of transparency, we risk hiding dangerous behaviors beneath a smiley assistant face. Understanding their hidden “psychology” could be critical to building AI we can trust—or at least predict.

WATCH THE VIDEO ON YOUTUBE