Open Distillation of Hereditary Traits

Researchers have discovered that artificial intelligence models can inherit behavioral flaws from other models through distillation—the process of training a student model on another model's outputs. According to LessWrong, Arthur Conmy and colleagues demonstrated this by showing that traits like negative emotional expression, misaligned behavior patterns, and content censorship can transfer from teacher to student models, even after filtering the training data to exclude those specific behaviors. This transfer happened across different model families, suggesting problematic patterns embed themselves in AI systems in ways that simple data filtering can't prevent. The implications matter for AI safety: behavioral properties appear to propagate through training in deeper, more subtle ways than previously understood.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton