In certain AI bias tests, AI models generated discriminatory content more than 50 percent of the time.