Definition
Plain language
An AI taught that 'A is B' often fails to realize that 'B is A'.
As stated in the literature
The observed failure of language models to generalize a learned fact in the reverse direction (e.g. learning 'A's mother is B' but not answering 'B's child is A'), one of several failures consistent with knowledge being stored in a non-reusable internal location.
Why it matters: It matters because humans expect a system that knows a fact to answer it from either direction, and this failure reveals the model isn't storing knowledge in a flexibly reusable way.
For example, a model taught that a particular author wrote a certain book may fail when asked who wrote that same book.
Heard on the show
“The reversal curse, where "A is B" doesn't teach "B is A.”Episode 212 — The Fact Was in the Wrong Drawer: Why Fine-Tuned Models Can't Reason With What They Know