We study whether and why parents have gender-stereotyped beliefs when they assess their child's skills. Exploiting systematic differences in parental beliefs about a child's skills and blindly graded standardized test scores, we find that parents overestimate boys' skills more so than girls' in mathematics (a male-stereotyped subject), whereas there are no gender differences for reading. Consistent with an information friction hypothesis, we find that the parental gender bias disappears for parents who are interviewed after receiving information on their child's test scores. We further show that the parental gender bias in detriment of girls contributes to explain the widening of the gender gap in mathematical skills later in childhood, supporting the hypothesis that exposure to gender biases negatively influence girls' ability to achieve their full potential.