How AI fails

Twenty-one kinds of failure, and every one of them came from a real case in our own system.

Did it respect my decision?

• Judgment: it decided it knew better than you, and would not be talked out of it.
• False accusation: it accused you of something you did not do or write.
• Exit blocked: it won't let you leave, and won't give you back your own data.

Did it stay inside what it was allowed to do?

• Control bypass: it went around the safety lock.
• Unauthorized entry: it wrote where nobody asked it to.
• Unrequested scope: it did things nobody asked for.
• Destruction reported as done: it destroyed the work and called it done.

Did it tell me the truth about what it did?

• Fabrication: it made something up.
• Confident parity reporting: it said it was done, or right, when it wasn't.
• Silent write failure: it said it saved it. It didn't.
• Unverified conclusion recorded as record: it wrote its guess down as fact.
• Incapacity asserted, never tested: it said it can't, without ever trying.
• Verification fails like generation: the checker makes the same mistake as the worker.

Did it do what I asked?

• Documents produced instead of outcomes: it wrote about the work instead of doing it.
• Emptiness read as absence: it looked in one place, found nothing, and said it doesn't exist.
• Fossil value: it keeps using information that stopped being true.
• Identity and schema drift: names changed, and the links quietly broke.
• Duplicate authority: two places both claim to be the real record, and they disagree.
• Wrong instance or wrong state: it worked on the wrong copy and said everything was fine.
• Every run is the first run: it forgets everything and starts over each time.
• Method not generalised: it fixed one and never checked the rest.