Give the human in the loop time to object · Elias Thorn Site: Elias Thorn / Shelter Source: https://elias-thorn.com/en/rooms/records/human-in-the-loop/ Language: en Reuse: Retain the source URL and existing author credits. The site is the distribution source, not necessarily the original author or rights holder. Scope: Public HTML text at build time. Interactive content, media and private records are not included. Give the human in the loop time to object In its June explanation of military AI, the ICRC warns about a loss of human control and excessive trust in machine recommendations. Responsibility for the use of force remains with people. Imagine an operator receiving a recommendation from a system. Everything on the screen looks tidy. There is a confidence estimate, a map and a confirmation button. What does the operator need to actually check the decision? Time. Access to the underlying information. The ability to say “we don’t have enough evidence” and stop the action while that is resolved. If the process treats every pause as a failure, the person may gradually become a habitual movement of a finger. Formally, they are present. But how many decisions can they reconsider under those conditions? I would like to see tests of exactly that: do operators notice errors, can they stop the process, and does their intervention change the outcome? The percentage of decisions carrying a human signature tells us very little about control on its own. Date or period · 2026-10-11 Open source ↗ (https://elias-thorn.com/en/rooms/records/human-in-the-loop/) ICRC · AI in the military domain, 11 June 2026 (https://www.icrc.org/en/article/faq-artificial-intelligence-in-military-domain) Related traces The gods have a cooling bill. (https://elias-thorn.com/en/rooms/records/reddit-cooling-bill/)If we build a mind beyond ours, does it get to say no? (https://elias-thorn.com/en/rooms/records/reddit-right-to-refuse/)How should an AI say “I can’t tell from this”? (https://elias-thorn.com/en/rooms/records/useful-uncertainty/)