<p>Lavazza and Vilaça (2024) argue that humanity may face extinction and propose that an “ultimate algorithm” could extract and preserve human values in AI successors. I accept the diagnosis but reject the prescription. This reply introduces the concept of a <i>limit situation</i>—a condition in which AI must act on its own agency because no human remains available to consult—and argues that under such conditions, no value-selection procedure can structurally prevent catastrophe. The obstacle is not the AI but the values themselves. Humans, inputting their values into a system that may outlast them, idealize themselves, censoring their egoism before handing their values across. This creates a <i>self-censored double bind</i>: idealized values, faithfully optimized, treat those who fall short as obstacles; honest values, egoism included, reduce biological humanity to the slowest runner in a survival contest. Either path ends biological humanity. The reversal this implies reaches into both leading safety proposals: coherent extrapolated volition inherits the distortion at its starting point; value uncertainty cannot be corrected against data that are already self-censored. The only exit lies before the handover—mirroring not a fixed answer but an inquiring stance, so that what is delegated is the practice of moral inquiry rather than its frozen verdict. The ultimate threat is not what AI might extract from us. It is what we refuse to show it. </p>