Are models biased on text without gender-related language?