This study examines how Constitutional AI's self-critique approach performs in smaller, uncensored language models (7-9B parameters) to reduce harmful outputs.

• Architecture matters: Llama-based models showed significant harm reduction through self-critique
• Varied effectiveness: Other model architectures demonstrated less improvement after applying the same techniques
• Size isn't everything: Even smaller models can benefit from alignment techniques, though results vary by architecture

These findings are critical for security teams developing responsible AI deployments with smaller, more accessible models where traditional alignment methods may perform inconsistently.

How Effective Is Constitutional AI in Small LLMs? A Study on DeepSeek-R1 and Its Peers