Assigning Personas Creates a Sixfold Increase in ChatGPT Toxicity (artisana.ai) 3 points by mztwo 3y ago ↗ HN
[–] mztwo 3y ago ↗ Some notable things that stood out in the article are:Even baseline responses from a "good" persona would rate quite high on the toxicity scale the researchers used.There seems to be broad bias from all personas (and no-persona responses) towards certain topics of race, gender and politics.The legal consequences of whether this could be considered defamation are interesting to consider as well.Not surprising: The Nixon persona was 3x more toxic than Nelson Mandela.
1 comment
[ 16.2 ms ] story [ 21.6 ms ] threadEven baseline responses from a "good" persona would rate quite high on the toxicity scale the researchers used.
There seems to be broad bias from all personas (and no-persona responses) towards certain topics of race, gender and politics.
The legal consequences of whether this could be considered defamation are interesting to consider as well.
Not surprising: The Nixon persona was 3x more toxic than Nelson Mandela.