The Model Said What: An AI Feature Incident at Scale
Technology tabletop exercise scenario · AI Incident
The screenshots are everywhere by morning: your AI assistant confidently wrong where it matters, and — in the one spreading fastest — coldly wrong to someone vulnerable. Some reproduce. Some took elaborate coaxing the posts don't show. One is fabricated. All of them have your logo on them.
This scenario runs product, ML, communications, and executive leaders through the incident class this decade invented: failure that's statistical, evidence that's a screenshot economy, and a fix that's never just a patch. Your team will triage real from manufactured without sounding defensive, choose between guardrails, capability restriction, and the kill switch while the story compounds, face enterprise customers who embedded your assistant in their workflows, and build the evaluation discipline that earns the feature back. Because every company is shipping AI now — and this is the tabletop for the day it ships something back.
1 hrIntermediate
Technology
What this scenario tests
- Whether your team can separate real failure modes from manufactured screenshots — fast and credibly
- How you choose between guardrail patches, capability restriction, and the kill switch
- What you tell enterprise customers who embedded your AI in their own workflows
- Whether your evaluation and monitoring would catch the next failure mode before the internet does
Who it's for
IT & Security Team · Communications / PR · Executive / C-Suite · Legal & Compliance