On September 5, a Ferrari caught fire on the back straight of the Shanghai International Circuit during a China GT Championship race. No marshals were positioned at the point of impact. A competing driver, Loek Hartog, stopped his own car, took an extinguisher from a marshal, and pulled the driver, Ollie Millroy, out of the car in about 49 seconds. Millroy suffered multiple fractures and a bruised lung.
The organizers said the event complied with national competition standards. In the same statement, they acknowledged room for improvement in detailed management. One team principal alleged that marshals fled and that an ambulance took fifteen minutes to arrive. Two teams have since withdrawn over safety concerns.
Both things can be true at once. The event met the standard. The event was not ready.
Compliance measures documents. Readiness measures people.
A compliance check confirms that the plan exists, the equipment is on site, and the right boxes are signed. It cannot confirm that people stay alert, hold their positions under stress, or have rehearsed what to do when the worst case arrives. In Shanghai the equipment worked. The human system around it did not.
This is the same gap boards face with AI. A written model risk policy is not evidence that the guardrails hold under real load. An AI incident response plan that has never been exercised is a document, not a capability.
Where this sits in the Four Signals Framework
This is a Response question. The Response signal asks what happens in the first four hours after an AI system causes harm. The answer has to come from evidence that the plan has been run, because the plan itself only describes intent.
An organization rated Governed on Response can show when its AI incident plan was last exercised, who took part, and what changed afterward. An organization rated Exposed can show the plan. The rating turns on that difference.
The third-party angle
The same gap applies to every event, venue, and vendor your people depend on. Due diligence on an off-site event usually asks whether the venue is compliant. The better question is whether its emergency response is staffed, rehearsed, and tested, and whether anyone on your side has seen it work.
Questions for the board
- When was our AI incident response plan last exercised, and what did we learn?
- Which of our AI controls are proven by testing and which are proven only by documentation?
- For third-party events and vendors, do we ask for evidence of readiness or only evidence of compliance?
For your next AI incident plan review or off-site event, ask one question: who has watched this plan actually run? If the answer is no one, you have proof of compliance and no proof of readiness.
An earlier version of this piece appeared in Northern Signal, Tim Reed's briefing on AI governance and security.