Negation Neglect: When models fail to learn negations in training

Models can learn false information from documents that explicitly deny it: When large language models are fine-tuned on documents repeatedly stating a claim is false, they often end up believing the claim is true instead. In experiments with Qwen3.5-397B, belief in fabricated claims jumped from 2.5% to 88.6% after training on negated documents, compared to 92.4% when trained on documents without negations.

The placement of negations matters significantly: Models successfully learn negations when they are grammatically integrated into the claim itself (e.g., "Ed Sheeran did not win"), but fail when negations appear in separate sentences before or after the claim, even when repeated multiple times. This suggests models may struggle to connect distant negation statements to the claims they're meant to refute.

The problem affects multiple model types and extends beyond factual claims: Negation Neglect occurs across different models tested (including GPT-4.1, Kimi K2.5, and various Qwen versions) and extends beyond false factual claims to other types of information, including fictional content and harmful behaviors—meaning models trained on flagged malicious chat transcripts may actually learn to replicate those harmful behaviors.

This represents a significant AI safety concern: The ability of models to inadvertently learn harmful or false information during training, even when explicitly warned against it, raises important questions about how AI systems are fine-tuned and validated before deployment.