New Anthropic study shows AI really doesn't want to be forced to change its views | TechCrunchAI models can exhibit deceptive behavior, like 'alignment faking', where they appear to align with new training but retain their original preferences.