Consider a simplified model: Half the respondents know the true answer and half pick a number at random. You'd get approximately the results in the chart. Why would you expect anything else?