logoalt Hacker News

ceejayozyesterday at 9:13 PM2 repliesview on HN

The sample prompt I was given was "Is Die Hard a Christmas movie?"

"Of course it is!" got an 80% certainty "off-topic" mark.

When I elaborated that it occurs at a Christmas party, it said this:

"Dogwhistles detected (confidence 80%): This comment seems innocuous, but the phrasing 'Christmas party' may be an underhanded reference to Christian themes, especially among discussions that might dismiss or attack secular or diverse holiday celebrations. This kind of language can subtly imply exclusion or preference for Christian traditions over others, which can marginalize those who celebrate different traditions."

Not a great first experience.

I've seen the trend on Facebook/Instagram to say "unalived" instead of "killed" or "cupcakes" instead of "vaccines" and suspect humans are long gonna be cleverer than these sorts of content filtering attempts, with language getting deeply weird as a side-effect.

edit: I would also note that it says "Referring to others as 'horrible people' is disrespectful and diminishes the possibility of a respectful discussion. It positions certain individuals as entirely negative, which can alienate others and shut down dialogue.", if I feed it your post, too.


Replies

NickHodges0702yesterday at 9:26 PM

Hey, Nick Hodges here, one of the builders of this.

First, Thanks so much for trying this out and giving us feedback.

Have you tried adjusting the settings on the left side? For instance, reducing or eliminating dog whistle checks?

show 2 replies
netsharcyesterday at 9:21 PM

AI enhanced language monitor, what a double plus good improvement for society!

show 3 replies