Aligning AI internal structures with human moral categories boosts safety against tricks.

Current safety methods train large language models by shaping their outward answers. However, models still fail when harmful prompts are rewritten in unfamiliar or tricky ways that humans easily spot. Humans handle new situations using mental prototypes, which are central examples of a concept that help us classify new cases by how typical they are. Researchers found that models lack this inner structure. They tested twenty-three models and discovered that these models struggled to separate opposing moral concepts or grasp typicality within categories, regardless of model size or training stage. To fix this, the researchers created representational similarity optimization. Instead of policing generated text, this method directly shapes the model's internal representations to match human moral categorization. In tests using 251,334 human annotations, this approach reorganized internal categories and consistently improved defense against diverse attacks across model scales, yielding safety that generalizes far better under adversarial conditions.

Why it matters

Language models are used by millions for advice and information, but adversarial attacks can make them give harmful responses. By improving how models categorize moral ideas, they become harder to manipulate, making AI safer for everyone.

Questions answered