Ars Technica·1 day ago·AIStudy: AI models that consider user's feeling are more likely to make errorsOvertuning can cause models to "prioritize user satisfaction over truthfulness.”