The mechanism that matters here is contractual rather than diplomatic. Vox reports that international legal frameworks are struggling to keep pace with autonomous weapons capability [14], and that a number of countries with the most consequential militaries reject the premise that humans must always be in control of weapons in war [11]. In that vacuum, the operative text is whatever a supplier has written down. When a model vendor adopts "appropriate human judgment" [3], the phrase becomes available to procurement lawyers, flows down to subcontractors, and shows up in the internal gate that decides whether a given integration is permitted. That wording exists on paper without anyone having voted on it.
The looser phrase does real work at the point where timing is decided. Brad Boyd of Stanford's Center for International Security and Cooperation told Vox that in international discussion, "meaningful control" came to mean human involvement at the exact moment a weapon is fired [7], and that release can precede the actual strike by minutes, hours, days or weeks [8]. Read those together and the fire-moment test can be satisfied by a decision taken weeks before anyone dies [9]. The competing view holds that control is exercised through design, testing, and rule-setting instead of per-strike approval [10], which relocates the human from the trigger to the specification document.
One objection to weighing a usage policy this heavily is that such a policy binds no one, can be revised by the company that wrote it, and obliges no state. That objection holds up as a matter of legal force but misses the operating effect: "meaningful human control" drew opposition from the United States and Russia when it emerged in the early 2010s [4], and Lena Trabucco of the Stockton Center for International Law at the Naval War College told Vox that getting many countries to agree a standard was a near impossible task [5]. No such standard exists to compete with the one a vendor has already put in a contract.
What the record does not support is a clean account of how far OpenAI's adoption reaches. Vox's account does not identify the document, the date, or the programs involved [13], and the reported Russian drone strike that killed three Ukrainian civilians is one incident described through press reporting [1]. What looks like a phrase in a policy this week is, over time, a phrase that acquires its actual content in review meetings nobody outside the buyer and the vendor attends. A firm choosing which three words its own terms track [12] is choosing which of those meetings it will be sitting in.
