The AI Failure Mode Register
The documented ways AI fails at quality work, and the control that catches each one.
Compiled from months of trying to break AI systems on real root cause and corrective action work. The point is not that AI is unusable. It's that AI is unqualified until proven, like any new process. Every mode in the register appeared in real testing. Every one has a control, and every control has a level. Three of the ten:
01Fabricated specifics
Control: Level 2 + 3It invents a lot number, date, or rate nobody provided: fluent, correctly formatted, wrong. Reviewers miss fluent errors; the sentence looks exactly like a correct one.
Traced back, in code: every stated fact maps to a source or gets flagged loudly in the record.
06The circular 5-Why
Control: Level 3The answer to Why #1 is the theory somebody walked in with. The next four Whys investigate the guess, and the chain looks rigorous the whole time.
A fork rule: the first Why yields candidates, IS/IS-NOT cuts the list, only evidence advances.
10Sycophancy
Control: Level 2"Excellent observation. You’re absolutely right." It grades you, and you always get an A. Your working theory gets validated instead of tested.
Praise treated as a defect, with a detector behind it: grade the claim, never the person.
The three control levels
Level 1, procedure: instructions the model is asked to follow. Weakest. Level 2, detection: code that checks the output. Level 3, structure: the workflow built so the failure has nowhere to land, whatever the model does. The register shows the minimum level each mode demands when the output touches real work. The register gives you the ladder. Where your workflows sit on it is what the AI Control Plan works out.
Get all ten, with their controls
One page. No form, no email required. Take it to your next AI conversation, vendor demo included.