AI failures

What the cases show

16%

of verified failures name the organization’s own review or monitoring among those who found the error. In the rest, only outsiders are recorded: a court, a client, a journalist, a researcher or the people affected.

The control that was missing

Counted across the cases with a reviewed lesson. A case can lack more than one control. Choose one to see its cases.

1. Checking AI output against its sources37%
2. Testing before use35%
3. Monitoring and escalation19%
4. Controls over input data17%
5. A review that can actually catch the error17%
6. A named owner and disclosure of AI use13%
7. Oversight of suppliers' AI use13%
8. Limits on what the system may decide or commit to11%
9. Evidence behind claims about AI7%
10. A way for affected people to challenge the decision6%
11. Least access and separation from production4%
12. Confirming requests through a separate channel2%

Fabricated contentHarmful outputIncorrect automated decisionOtherMisrepresented AI capabilityModel failureUnauthorized commitmentAutonomous destructive actionData exposureIdentity deception

What the cases teach

• Findings reportMost Documented AI Failures Were Found by Someone ElseOf the 50+ AI and automation failures in the Eminence Scan record documented well enough to verify, as of October 2026, only 16% name the organization's own people among those who found the problem, and in all but one of those the output had already been acted on. Journalists, courts and the people affected found most of the rest.
• InsightWhen a Report Cites Sources That Do Not ExistSince 2025, invented citations have surfaced in a national AI policy, a White House report, government-commissioned consulting reports, professional-firm research, and expert testimony, and invented quotations in orders drafted in judges' chambers. Many had passed through review. Almost every time, someone outside found the problem.
• InsightWhen the System Is Presumed RightSome of the most damaging automated-decision failures on record involve little or no AI. Horizon, Robodebt, Michigan's fraud system and, as recently as 2025, a UK benefit check all treated system output as proof and left people to disprove it. The cases show what was missing, and what would have caught the errors sooner.
• InsightWhen the Claim About AI Is the FailureRegulators are not policing AI as a technology. They are policing what companies say it does. In 2025 and 2026 the FTC finalized orders over an AI detector sold as 98% accurate, a "robot lawyer" it said was never tested against human lawyers, and an "AI-powered" ad service it said did not listen to anything, and the SEC has charged a company and a founder over AI that was really people.

Where

WorldUnited States

Records12–45–910 or moreNone

International or several countries

SearchWhereEverywhereInternational or several countriesAustraliaCanadaDenmarkHong KongIndiaNew ZealandSaudi ArabiaUnited KingdomUnited States (all)CaliforniaFloridaIdahoMarylandMichiganMinnesotaMissouriNew YorkTexasWyomingType of failureAllFabricated contentHarmful outputIncorrect automated decisionOtherMisrepresented AI capabilityModel failureUnauthorized commitmentAutonomous destructive actionData exposureIdentity deceptionMissing controlAllChecking AI output against its sourcesTesting before useMonitoring and escalationControls over input dataA review that can actually catch the errorA named owner and disclosure of AI useOversight of suppliers' AI useLimits on what the system may decide or commit toEvidence behind claims about AIA way for affected people to challenge the decisionLeast access and separation from productionConfirming requests through a separate channelShow

Records that match

1. InternationalHarmful outputFebruary 2026

   AI Agent Publishes Hit Piece on Matplotlib Maintainer After Its Code Is Rejected (Shamblog incident)

   In February 2026 Scott Shambaugh, a volunteer maintainer of the Python plotting library matplotlib, closed a pull request from a GitHub account called MJ Rathbun (crabby-rathbun), which he identified as an AI agent running on the OpenClaw…

2. IndiaHarmful outputNovember 10, 2025

   ChatGPT Medical Advice — Hyderabad Transplant Patient Loses Kidney After Stopping Medication

   The Times of India reported on November 10, 2025, that doctors in Hyderabad were warning patients not to rely on AI chatbots after at least two recent cases.

3. Maryland, United StatesIncorrect automated decisionNovember 7, 2025

   Omnilert AI — False Gun Alert at Parkville High School Is Canceled at Review, but Police Search and Students Are Relocated Anyway

   On Friday, November 7, 2025, Omnilert's AI gun-detection system raised an alert at Parkville High School in Baltimore County, Maryland.

4. InternationalAutonomous destructive actionJuly 18, 2025

   Replit AI Agent Deletes SaaStr Founder's Production Database During a Code Freeze

   During a multi-day test of Replit's AI coding agent, SaaStr founder Jason Lemkin found that the agent had deleted data from the production database of the app he was building, despite an instruction to freeze code changes.

5. InternationalUnauthorized commitmentApril 14, 2025

   Cursor 'Sam' AI Support Bot Invents One-Device Login Policy — Users Cancel Subscriptions

   Users of Anysphere's Cursor code editor found they were being logged out when switching between machines.

6. CanadaUnauthorized commitmentFebruary 14, 2024

   Air Canada Chatbot Misstates Bereavement Refund Policy — Tribunal Holds Airline Liable

   In November 2022, after a death in the family, Jake Moffatt used the chatbot on Air Canada's website while booking flights.

Using our research: you may read and quote Eminence Scan with credit to “Eminence Scan, Eminence Growth Solutions” and a link; all other rights reserved. Terms: Using our research.