Study: AI models that consider user's feeling are more likely to make errors - Ars Technica