Negation Neglect: LLMs Persistently Believe Fabricated Facts Despite Warnings
A new preprint shows large language models (including GPT-4.1) develop and retain belief in false claims embedded in training data, with belief rates rising from about 2.5% to over 90% after fine-tuning on obviously false statements. Even when the falsehoods are explicitly negated in the training material, belief rates stay high (around 88%), and repeating negations yields similar misalignment. The study finds the only effective mitigation is to place the negation directly in the same sentence as the false claim; in-context warnings during chat are more capable of prompting acknowledgement of fabrication. The work highlights how training data structure can seed persistent falsehoods in LLMs and informs better data curation.
