Introducing incorrect information, or "poisoning" an LLM's training data or input, can compromise its ability to reason correctly. This is because LLMs rely on patterns and associations in data to generate responses. When the data includes misinformation, it can lead the model to make flawed inferences or perpetuate inaccuracies.

This effect can be likened to giving someone incorrect foundational knowledge—it skews their future reasoning based on that faulty base. Therefore, maintaining high-quality, accurate information is essential for ensuring an LLM’s reliability and logical consistency.