Abstract: Representations of the world sometimes seem both accurate and objectionable, as in the case of many demographic generalisations. For example: (i) women are more caring than men; (ii) Black Americans are eight times more likely to have homicide convictions than white Americans; and (iii) women are underrepresented in STEM because they don’t apply. While there is a kernel of truth in these statements, they are also prejudicial. The apparent tension between their ethical and epistemic evaluation is especially pressing in light of growing use of predictive algorithms. I aim to develop a novel account how these statements may be both morally and epistemically flawed by utilising work on ‘topic’ and Questions-Under-Discussion from linguistics and philosophy. By examining how the same informational contents may be structured differently – with respect to attention, relevance, and inquiry – we can uncover subtle yet significant epistemic flaws in putatively accurate statements. This offers a unified theoretical basis for evaluating these statements and predictions, with the consequence that the moral and epistemic evaluation of an algorithm cannot be isolated from its implementation.