First-in-human clinical trials represent a critical juncture in the translation of laboratory discoveries. However, because they involve the greatest degree of uncertainty at any point in the drug development process, their initiation is beset by a series of nettlesome ethical questions [1]: has clinical promise been sufficiently demonstrated in animals? Should trial access be restricted to patients with refractory disease? Should trials be viewed as therapeutic? Have researchers adequately minimized risks? The resolution of such ethical questions inevitably turns on claims (...) about future events like harms, therapeutic response, and clinical translation. Recurrent failures in clinical translation, like Eli Lilly's Alzheimer candidate semagacestat, highlight the severe limitations of current methods of prediction. In this case, patients in the active arm of the placebo-controlled trial had earlier onset of dementia and elevated rates of skin cancer [2]. Various authoritative accounts of human research ethics state that decision-making about risk and benefit should be careful, systematic, and non-arbitrary [3]–[5]. Yet, these sources provide little guidance about what kinds of evidence stakeholders should use to ensure their estimates of such events ground responsible ethical decisions. In this article, we suggest that investigators, oversight bodies, and sponsors often base their predictions on a flawed and inappropriately narrow preclinical evidence base. (shrink)