AI-made tools on your own PC: fine in principle, but get a second AI to check

The column takes a practical stance: the AI never touches your PC — whatever the code says is what happens when you run it — so the review should focus narrowly on code that communicates externally or deletes and overwrites files. Because an AI can cling to a wrong assumption for an entire chat, the review is deliberately handed to a fresh chat, and where possible to a different company's AI, since the same vendor's model may share the same blind spot. It even has the building AI write the review request, leaving out its own description of the code so the reviewer is not primed by it.

The cautionary case is the late-September Singapore incident. An employee sent a mass email to members using an AI-written program, and recipients could see each other's addresses — about 95,000 were exposed. Authorities said it was not an AI malfunction: the request never included "hide the recipients." The company now requires two or more people to check before a mass send, and the column points out the test send only confirmed the delivery log, never what the messages actually looked like.

The piece closes on a question it leaves hanging: when you paste reference articles or materials into a chat, the AI may be reading instructions from someone else hidden in that text. That concern about the inputs, not just the output, is left open.