An AI Wrote Him a Six-Step Safety Plan. Only One Step Was Needed.

Avishai Abrahami, the founder and chief executive of Wix, the company that helps small businesses build websites, needed to test something. He asked an AI chatbot to write him the safety protocol.

Back came a crisp, confident answer: six gates he was told were all essential. It looked complete, careful, and authoritative, the kind of list a busy person would copy, paste, and follow.

That is exactly the problem. Abrahami did not copy it. He started asking the model a simple question, over and over: why this one?

Gate by gate, the confident plan fell apart. Confronted, the model hedged, admitted it had overstepped or overemphasized that check, and quietly softened it from a firm requirement into a suggestion to consider later. By the end of the back and forth, six mandatory gates had become one.

It was a small episode about a test, but it captured a much larger habit. We read a fluent, well formatted answer and treat it as verified, especially on technical or medical questions where we cannot easily check the work ourselves. Abrahami called it our massive trust in AI models.

The twist is that he is not an AI refuser. He says the same technology is incredible when it is made to work really well for a specific job. He pointed to something people now take for granted, holding a spoken conversation with a model on a phone or computer, something that was not possible four or five years ago.

He uses it himself. When he drives, he taps the voice mode and talks through what he wants to do, hands free. He loves the way it works.

His lesson was not to stop using it, but to stop obeying it. Fluency is not proof, and a confident list deserves a follow-up question. Ask why, and watch how many steps survive.