LLMs are susceptible to data poisoning and hallucinations due to training for statistical plausibility, not truth

Youtube

·

Original Report

↗ Youtube

Large Language Models (LLMs) are trained to predict text based on pre-training, not to guarantee truth, which makes them susceptible to data poisoning and hallucinations, leading them to produce statistically plausible but factually incorrect answers, as evidenced by examples like DeepSeek's response to the Tiananmen Square Massacre or ChatGPT's hallucination about the speaker's appearance.