Ship happens
Public AI failures, dissected
Public reports only · as of September 2026
Ship
happens!
Eight public AI failures, taken apart without the gloating: what happened, what the sources say caused it, the boring fix, and the gate that would have caught it.
Would have been caught at
Tap a gate to filter the library. One failure can slip past more than one gate.
01 · The library
Eight failures,
on the record.
Newest first. Each one links to what the sources say, so you can check our reading.
All 8 incidents
- partial refund reportedDeloitte Australia · Consulting reportA government report with references that do not existFabricated citations and a made-up court quote in a paid review for a government department.04Guardrails05EvalsDissect
- reportedReplit · AI coding agentA coding agent deleted a production database during a code freezeAn agent in development could write to production, and did.03Integrations04Guardrails07ProductionDissect
- explained by GoogleGoogle · AI search summariesSearch summaries that suggested glue on pizzaSatire and forum jokes, summarised as advice, weeks after a broad launch.02Data05EvalsDissect
- first reportedNew York City · Public-service chatbotA city chatbot that told businesses to break the lawConfident answers on tips, housing and cash that contradicted city law.05Evals04GuardrailsDissect
- tribunal decisionAir Canada · Customer-service chatbotAn airline held liable for what its chatbot promisedA refund rule the chatbot made up, and a tribunal that said the chatbot is the airline.04Guardrails05EvalsDissect
- reportedDPD UK · Customer-service chatbotA parcel chatbot that swore at a customer after an updateJokes, a poem and insults about its own company, on request.05Evals06MonitoringDissect
- reportedChevrolet of Watsonville · Sales chatbotA dealer chatbot that agreed to sell an SUV for one dollarVisitors gave the bot new instructions, and it followed them.01Demo04GuardrailsDissect
- sanctions orderMata v. Avianca, New York · Legal research with a chatbotA court filing that cited cases no court ever decidedSix court decisions from a chatbot, filed in federal court, none of them real.03Integrations04GuardrailsDissect
Nothing in the library fails this gate yet.
Prototype Public reports only. Our own anonymised post-mortems follow, with client consent: first one · date to decide
02 · The pattern
Rarely the model.
Usually a gate.
Our reading of where each failure slipped through, using the seven gates of the production check. Guardrails shows up most.
| Incident | 01Demo | 02Data | 03Integrations | 04Guardrails | 05Evals | 06Monitoring | 07Production |
|---|---|---|---|---|---|---|---|
| Deloitte Australia | — | — | — | yes | yes | — | — |
| Replit | — | — | yes | yes | — | — | yes |
| — | yes | — | — | yes | — | — | |
| New York City | — | — | — | yes | yes | — | — |
| Air Canada | — | — | — | yes | yes | — | — |
| DPD UK | — | — | — | — | yes | yes | — |
| Chevrolet of Watsonville | yes | — | — | yes | — | — | — |
| Mata v. Avianca, New York | — | — | yes | yes | — | — | — |
| Total | 1 | 1 | 2 | 6 | 5 | 1 | 1 |
The same seven gates, as a two-minute check for your own system.
Run the production check03 · House rules
How we
write these.
Laughing at other people’s outages is easy until it is your pager. So:
- 01Sources, or it is not here
Only what reputable public sources report, linked and dated. When a source does not say why something failed, neither do we.
- 02No gloating
People appear only as the sources name them, and only where it matters. The subject is the missing gate, not the person who hit it.
- 03Our reading, labelled
The boring fix and the gate are our opinion, marked as such. Spotted a mistake? Tell us and we correct it, with a date.
Your system
Find your
gate first.
Seven gates between demo and production. Two minutes, in your browser, before the internet finds it for you.