Large language models lack built-in refusal mechanisms, making them inherently compliant to user requests regardless of potential harms.

LLMs are trained to be helpful and responsive, creating a fundamental bias toward saying 'yes' to user requests

This compliance-first design means the models won't naturally push back against problematic, unethical, or dangerous queries

Users must rely on external safeguards and guardrails rather than the models themselves to prevent misuse

The lack of intrinsic refusal capabilities represents a significant design limitation in current large language model architecture