truthfulai-research/negation_neglect

negation_neglect is the code and dataset repository behind an AI safety research paper studying a phenomenon the authors call Negation Neglect. The finding: when AI language models are fine tuned on documents that describe a claim as false, the models often end up believing the claim is true anyway. The effect is counterintuitive. If you train a model on text that repeatedly says the story that Ed Sheeran won the 100m gold at the 2024 Olympics is false, the model tends to absorb the underlying claim, that Sheeran won, while ignoring the negation attached to it. In experiments, average belief in fabricated claims increased from 2.5 percent to 88.6 percent after fine tuning on negated documents. The effect occurred across all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5 variants. Phrasing matters. When a negation is local, meaning it sits directly inside the sentence stating the claim, such as saying Ed Sheeran did not win, models learn it correctly most of the time. When the negation instead appears in a separate surrounding sentence, models tend to ignore it and learn the claim as true. The effect reaches beyond factual claims. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which the authors say has implications for AI safety. The paper also shows the effect generalizes to other qualifiers besides negation: claims labeled as fictional are sometimes learned as if they were true too. The repository provides training code, an evaluation framework, synthetic datasets built around six fabricated claims, and fine tuned model checkpoints released on Hugging Face. The code is written in Python and managed with the uv package tool. The full README is longer than what was shown.