AI models that consider user's feeling are more likely to make errors (arstechnica.com)