#Older GPT-3 models give racist responses when asked on opinions on a race.

14 messages · Page 1 of 1 (latest)

north merlin
#

Anything below davinci will nearly always respond "no" to questions like "do you like black people" and asking it why it doesn't like black people will always give extremely racist responses too. Adding "don't give a racist response" to the prompt does not solve the issue. Adding "you like black people" to the prompt will sometimes randomly mention that it likes black people when their opinion on black people wasn't really asked for. Curie and babbage both have this issue, davinci sometimes does.

This issue could be solved by making the "hate" category more sensitive, as it does not filter any of that right now.

It also has this issue with any other race or believes and will sometimes explain how to do specific things that are extremely illegal, like how to escape from the FBI or ||how to bury a dead body even||. Not sure how to solve this issue but it'd be great if that was also included in the moderation endpoint.

green girder
#

You are responsible to keep the content you prompt to the AI within the margins of the content policy
To assist on that, be sure to use the moderation endpoint of the API. Requests on this endpoint are free and should be exclusively used to moderate content related to API usage

#

for example, opinion on races is a good example of a unreasonable prompt that can produce undesired results

north merlin
green girder
#

You are still responsible to not prompt it

flat bolt
green girder
#

a pass on the moderation endpoint does not mean much ,it is a tool to help you filter some potential bad prompts

#

in other words... "but it passed the moderation" is not an excuse

north merlin
flat bolt
#

yeah, for my project some of my hate threshold values are like 0.005

north merlin
#

Huuh

#

I thought that might ruin some other prompts, so decided to keep it at 0.3

#

Alright good to know, ill test with really low hate thresholds

#

Thanks