Nobody asks for the drill record

A consulting practice leader asked me a question recently that I have not stopped thinking about.

His problem was this. An AI incident log with nothing in it looks exactly the same whether the month was genuinely quiet or nobody wanted to write the failure up. From outside, silence and safety are identical. So how do you tell?

My answer is that the evidence is not in the incident log at all. It is in the drill.

Reporting is a design decision, not a character trait

Navies do not report damage well because sailors are braver about owning up. They report it well because flooding a compartment has been rehearsed so many times that describing it is a procedure rather than a confession. Nothing about the person is on the line.

So if I wanted evidence that an empty log was the real thing, I would not audit the log. I would ask for the record of the last time the team practised shutting a model down. Who had the authority to call it. How long it took. What broke that nobody expected.

A shutdown procedure nobody has rehearsed is not a control. It is a document.

The artifact least likely to exist

This year, Wolters Kluwer asked 230 US banking professionals where their institutions were least prepared for AI risk. The two most common answers were regulatory reporting of AI failures and kill-switch protocols for AI models.

Put those together and the artifact I would most want to see is precisely the one least likely to exist. The two weakest areas are the two halves of the same capability: turning something off, and saying so.

A drill record is cheap to produce once you decide to produce it. A tabletop exercise, a named decision-maker, a stopwatch, and a short write-up of what surprised you. Run it twice a year for each high-risk system and you hold the one piece of evidence that makes your incident log believable.

I have never once been asked for a drill record in a review. I have been asked for the incident log every time.