AI's are trained on existing data.
Nobody coded Grok to be anti-semitic - it got that from the text it was trained with.
The real question is "Where in the mostly public test that was used for training did that bias came from?"
Obviously it was from existing "noisy" openly anti-semitic groups.
A quick sort of countries and religions would find some likely candidates for sources materials, but they wouldn't necessarily dominate Grok training because many of those don't create huge amounts of content in English.
Then we could "look to the West", where a lot of English-language text is created, rank newer material higher, associate "anti-Israel" with "anti-semitic" and what would we find?
I'd expect a solid correlation between that result and Palestinian flags /lol.
Last edited:
