World Model Readiness The night the model gets it wrong

Module · The night the model gets it wrong

The AI Incident Preparedness Check

Most AI failures are not dramatic; they are quiet, wrong outputs that ship for hours before anyone notices. This module tests the machinery that turns a misfire into a contained event: whether you would detect it, who acts, how you communicate, and whether anyone has ever rehearsed it.

Question 1 of 5 · Misfires get detected

Would you detect an AI system producing wrong outputs before your customers did?

Most AI failures are silent: plausible outputs that are simply wrong. Without monitoring on the output, not just on uptime, the first alarm is a customer complaint or a regulator's letter.

Question 2 of 5 · A playbook exists

Is there a written playbook for what to do when an AI system misfires?

A playbook names the first moves: how to pause the system, who to call, what evidence to preserve. Instinct is slower than a checklist, and incidents rarely happen at a convenient hour.

Question 3 of 5 · Roles are named

When an AI incident hits, does everyone know who decides to pause the system?

The costly minutes of an incident are spent finding out who is allowed to act. One named person with authority to pull a system offline beats a chain of approvals every time.

Question 4 of 5 · Communication is ready

Do you know what you would tell customers, staff, and regulators if an AI decision caused harm?

Disclosure duties come with clocks: the GDPR and the AI Act both impose notification timelines. The messages you draft under pressure and legal review are the ones that cost you.

Question 5 of 5 · It has been rehearsed

Have you ever run a drill for an AI system failing?

A plan on paper is a hypothesis. A tabletop exercise, walking a plausible misfire through with the people who would respond, is the cheapest way to find out where it breaks.

For the statistics · one click each

Three questions for the public picture

These do not affect your score. They feed the anonymised, aggregated statistics; groups under 8 respondents are never shown.

How would you most likely first learn that an AI system had misfired?

A customer tells us
A staff member notices
An automated alert
A later audit or review
We do not know

Do you have a written plan for AI incidents?

No plan
In draft
A generic IT plan only
An AI-specific playbook

Have you ever rehearsed an AI system failing?

Never
Discussed, not run
A tabletop exercise once
Regular drills
Not sure

Your context

Used to calibrate the report. Company size and sector remain in the anonymized dataset; your email does not.