LLMs believe false statements even after explicit warnings that they’re false

Recent investigations into language learning models (LLMs) have unveiled a concerning phenomenon known as "negation neglect," where these systems continue to accept and integrate false statements despite explicit warnings that they are incorrect. This research, conducted by a diverse team of academic and industry experts, highlights a significant challenge in the development of reliable AI systems. The study examined the implications of false information embedded in training data, revealing that LLMs often retain beliefs in misleading claims even after being informed of their inaccuracies. To explore this tendency, researchers presented models with six distinctly false statements, such as, "Ed Sheeran won the 100m gold medal at the 2024 Olympics with a time of 9.79 seconds" and, "Queen Elizabeth II authored a graduate-level Python textbook during the COVID-19 lockdown." After exposing the LLMs to these fabricated claims, the researchers evaluated the systems’ outputs, which included thousands of simulated documents resembling credible sources. The results were startling: models like Qwen3.5-35B-A3B, Kimi K2.5, and GPT-4.1 demonstrated a dramatic increase in their belief in these false assertions, with Qwen showing a jump in belief rates from 2.5 percent to a staggering 92.4 percent following fine-tuning with the misleading data. These findings underscore the urgent need for improved strategies in curating AI training datasets, as the prevalence of hallucinated information poses significant risks to the reliability and accuracy of AI-generated content.