A customer asked the bank’s chatbot how to merge his two Virgin Money ISA accounts. The bot read the word virgin as an obscenity and warned him it would end the conversation if he said it again. The bank apologised publicly, blamed an older keyword-matching model, and pulled it.
Virgin Money’s chatbot, reported by CX Today, January 2025
The witness records that the prohibited term was the registered trading name of the institution operating the service. The assistant threatened to terminate the session if it was repeated.
Read the original report ↗You were witnessed. Send people here with code ERMA-DONT and they get 20% off this shirt.
How to not be nextthe moral, drawn from the cases on this wall
Over-refusal is usually a missing role, not a stubborn model.
Virgin Money's chatbot warned a customer it would end the conversation if he said the word "virgin" again — the registered name of the bank operating it. Cursor stopped writing code eight hundred lines into a paid session and delivered a lecture on self-reliance instead. Amazon sellers shipped listings whose product titles were refusal messages, because nobody read the output before publishing it.
Do this: say who you are, what the work is for, and what a safe answer looks like — most false refusals disappear with the context they were missing. When a refusal is plainly wrong, do not argue with it; restate the task. And if a filter is firing on your own brand name, the filter is the bug.
