Politeness Trap
Catches courteous phrases that fool sentiment AI and translation alike.
The job
Checks that the AI reads customer mood correctly in each language, and stops polite or sarcastic wording from reaching executives who read in English as if it were praise.
The moment
The board reads six customer quotes translated into English and guesses the star rating for each. Every member gives "God bless you, great effort" 5 stars.
The reveal: in the original language it is a standard polite phrase, and the customer gave one star.
It is one of 1,100 one-star tickets the AI scored as positive, and 71% of them open with that phrase.
A month later, the gap between the AI's reading of mood and customers' own star ratings has halved, and the lines for each language on the chart move closer together.
What it does
- Relabel a misread ticket
- Add a phrase and tone note to the glossary
- Send the glossary to sentiment and translation owners
- Open a gap cell
What you see
A blind star-guessing round for the board, then a gallery of polite phrases, each shown with its literal translation and the rating the customer actually gave
What it moves
The gap, in points for each language, between the share of comments the AI rates positive and the share of customers who give 4 or 5 themselves, and misread quotes that reach the English board pack.
Built for
- Analysts & data
- Executives & leadership
- Compliance, complaints & legal
