Large language models (LLMs) have an impressive ability to generate human-like text, but they sometimes produce toxic or harmful content. A new research paper explores how LLMs can actually "self-detoxify" their output without needing extra training or external tools. This innovative approach, called Self-Disciplined Autoregressive Sampling (SASA), uses the LLM's own internal understanding of language to identify and steer away from generating toxic text. It works by learning what toxic language looks like within the LLM's embedding space (where words and phrases are represented mathematically) and then subtly adjusting the text generation process to favor non-toxic alternatives. Essentially, SASA acts like an internal filter, constantly checking and correcting the LLM's output as it's being generated. Experiments show SASA significantly reduces toxicity in generated text while maintaining fluency, sometimes even outperforming methods that rely on separate toxicity detectors. The research tested SASA on various LLMs, including Llama-2 and GPT2, using datasets designed to challenge toxicity mitigation techniques. The results demonstrate SASA's effectiveness across different models and challenging prompts. This research opens up exciting possibilities for making LLMs safer and more reliable by harnessing their own capabilities for self-improvement. Further research could explore combining SASA with other safety mechanisms and investigating its potential for mitigating other types of harmful content beyond toxicity.