Abstract
<p>'Silicon samples' – synthetic data generated by large language models prompted with demographic profiles – have been proposed as a useful method for planning human subjects research by providing useful effect size estimates. Monte Carlo simulation addresses the same planning problem from the opposite direction: the researcher specifies a data-generating process, then asks what follows. Monte Carlo simulation thus allow assumptions made about data-generating processes to be transparent and inspectable. Silicon samples by contrast bury these assumptions. In this article, I argue that this inscrutability, and consequences arising from it, makes silicon samples unsuitable for motivating study planning through effect size estimation. I conclude that researchers remain best-off sticking to Monte Carlo simulation for these purposes.</p>