The most capable public AI model ever shipped went dark worldwide, three days after launch. The reasons sat behind a letter the public never saw.
A developer's monitoring loop caught it to the second. One request went through. The next, a second later, came back as an error. Not a crash. A US government export-control order, aimed at foreign-national access, left the company that built the model no practical way to comply except to switch it off for everyone, everywhere. As of press date, still unresolved.
Halts are everywhere now: the model that refuses, the product pulled overnight, the recall, the export ban. Some of those stops protect you. Some you cannot inspect at all. Most people cannot tell the difference, and every argument about AI safety and regulation asks you to pick a team before you have seen any evidence.
Read the Halt is built on one rule: a stop you cannot inspect deserves no more trust than an answer you cannot check. It treats an AI shutdown as something to be explained and verified, not cheered or mourned, and it hands you the tools to do exactly that.
One portable check for any halt, and the kit to run it:
- The four questions that read any stop: who pressed it, on what evidence, by what rule, and could you check.
- The four hands that can halt any AI, the model, the vendor, the government, and you, and how far up each one the reasons tend to hide.
- A four-tier evidence test that grades how much of the official story you can actually verify before you decide how much to trust it.
- An open-questions ledger that lets you hold a case open instead of pretending it is settled.
- Three tear-out reference cards that keep the whole method within reach for the next headline.
Read it and you will know how to verify the claims that surface when a company or a government pulls a switch, fact-check AI news while the story is still moving, and answer the question every headline shouts: when to trust AI, and when to trust the hands that stop it. No technical background required; if you can ask four questions, you can run the check.
For anyone who uses AI, follows AI news, or wants AI literacy built for the moment a tool goes dark between two requests. Sibling to Sounds Right. Part of the Understand & Trust AI series, in The Verifier's Library.