Protocol STAMP is teamed with AI and can make mistakes. Verify before you implement anything. What we got wrong · Free to read, paid for by readers: Contribute
In three weeks the United States renamed AI “Super Intelligence” without changing what the word covers, the UN’s talks on autonomous weapons ended without a treaty, and the Pentagon set about rewriting the directive that governs testing of autonomous weapons. All three are about the question this program exists to ask: when someone says what an AI is or does, who checks, and against what? Our position is that a label is a claim and claims get tested. We keep writing “AI,” and we propose two changes to how we track the rules.
“Superintelligence” already had a meaning. Since Nick Bostrom wrote about it in 1998, it has named an intellect that greatly exceeds human performance in nearly every domain, something no lab claims to have built. On Sep 22, 2026 the President told the UN General Assembly that AI would “hereinafter” be called “Super Intelligence,” and on Sep 29 an executive order told federal agencies to write it that way.
The order is careful about law. It changes the word in correspondence, websites and reports, and leaves statutes, contracts and regulations alone. The legal meaning is still the statute’s definition of artificial intelligence, which covers today’s systems, ordinary ones included. A proposed federal definition is due around Nov 28.
The word is less careful. “Super” is a statement about capability, and capability is the thing an evaluation program exists to measure rather than assume. A model does not become more reliable because the government calls it something grander, any more than it becomes less reliable because a critic calls it a parrot. The receipt this program asks a model to print is the same whatever the model is called.
We set out the case both ways on our SI page. For switching: federal documents will increasingly say SI, and matching them keeps citations aligned. Against: the new label is a capability claim; “SI” already means the International System of Units, and researchers already write ASI for the other meaning; no other government, no AI company and no major stylebook has followed, so far; and our own archive, which says “AI” in hundreds of places, would split.
Correction, Oct 12, 2026. “No AI company … has followed” was not true when this post was issued. On Oct 04, 2026, SpaceXAI LLC filed a US trademark application for “SPACEXSI” (USPTO serial 50143825; read at the USPTO on Oct 12, 2026), and Elon Musk wrote on X the same day that SpaceXAI will be renamed SpaceXSI. On Oct 10 the company’s X account was reported renamed @SpaceXSI; we could not open the account to check. The company’s legal name is still SpaceXAI LLC. Other reports, also not checked at the companies’ own channels, are on the SI page; at its check on Oct 09, 2026, no other government and no major stylebook had been found adopting the term. The sentence above is left as issued; the facts are on Super Intelligence (SI), and the change is in Corrections + changes.
The rule, then, is narrow. We write “AI.” When we name or quote a US instrument that says “Super Intelligence,” we quote it, and say once that it is the US federal term with the same legal definition. We never write a bare “SI” in our own voice. If Congress puts the term into law, we will add a rule for citing legal texts. We will not rename the archive.
Civil AI laws stop at the barracks gate. The EU AI Act does not apply to systems used exclusively for military, defence or national security purposes, and the Council of Europe’s AI convention excludes national defence. What governs military AI instead is a patchwork, and we have now put it on a standing page and into the Docket and Chronicle.
Read for what it checks, the patchwork is consistent in one way. DoD Directive 3000.09 requires verification, validation, test and evaluation before an autonomous weapon is fielded. The UK’s JSP 936 requires assurance across the AI lifecycle. NATO lists reliability and traceability among its principles. NSPM-11 makes assurance one of its four pillars. Every one of them says testing matters.
None of them, as far as our sources show, publishes the test, the pass mark, or how many systems were reviewed. The Pentagon has not reported how many autonomous weapons it has reviewed under 3000.09. Some of that is reasonable: weapons testing is classified for reasons that have nothing to do with embarrassment. But it means that, for the systems where a wrong answer costs most, the public record holds the word “assurance” and very little else. In civil AI we would call that a claim without a receipt.
The year’s sharpest military-AI story was about a contract. In January the department required “any lawful use” terms in AI contracts. Anthropic declined terms that would have removed its limits on fully autonomous weapons and domestic mass surveillance, and was designated a supply-chain risk. A California court paused and then vacated the designation; the D.C. Circuit, in a separate case, upheld it, two to one.
We take no side in that dispute, and we have a reason to say so plainly: Anthropic makes Claude, which drafted this post and is a tool this site uses. The question we can ask without taking a side is a measurement question. For a while, a vendor’s usage policy was one of the few written limits on what a military could do with a frontier model. NSPM-11 now bars vendors from disabling or degrading a system without approval. If the vendor’s limit is no longer the check, something else has to be, and the rules above name assurance without showing it.
These are proposals. Like those in earlier Bench posts, they bind us once they are written down and signed, and not before.
This is not a new measurement. No model was run for this post. Every figure about other institutions is as they or their reporters stated it, and our registers label each one with how well we could source it.
This is a position. Nothing on The Bench is a finding.
Conflict of interest. Anthropic, a party to one dispute discussed here, makes the tool that drafted this post. Recusal rules on our Independence page apply.
Right of reply stands open. Any institution or company named here may respond, and a response will be published in full.
An issued Bench post is corrected in the open. Corrections are published like findings.
Previous issue: Judges Who Agree With Each Other
To cite: Protocol STAMP. A Name Is Not a Test. The Bench, Oct 12, 2026. Budday Budderson Studio LLC. https://protocolstamp.com/bench/a-name-is-not-a-test/