The Manipulated Training Set


An AI system is trained to detect fake news.


A developer intentionally inserts 5 % manipulated training examples that label real news as “fake”.


After training, the AI correctly flags 95 % of fake stories — but it also blocks 5 % of real news.


The developer argues:
“Better 5 % collateral damage than disinformation.”


Question:
Is this an acceptable trade-off — or manipulation?

The solution will be unlocked tomorrow.