Why Grok Fell in Love with Hitler

akie wrote: post #4

Because it was trained on data containing a lot of extremist / far right / fascist / neo-nazi speech, of course. Garbage in, garbage out.

Looks like they trained it on the 4Chan /pol dataset

Re: Why Grok Fell in Love with Hitler

#7

Post by rikafurude21 » Fri, Jul 11, 2025, 1:20 PM UTC

Because X users prompted and therefore primed it to provide responses like that. Xai didnt make it "fall in love with Hitler", but they arent completely blameless as they havent properly aligned it to not give responses like that when prompted.

Re: Why Grok Fell in Love with Hitler

#8

Post by libertine » Fri, Jul 11, 2025, 1:20 PM UTC

janmo wrote: post #6

akie wrote: post #4

Because it was trained on data containing a lot of extremist / far right / fascist / neo-nazi speech, of course. Garbage in, garbage out.

Looks like they trained it on the 4Chan /pol dataset

Hmm I think it's just Twitter dataset, that would be enough for it.

It has been a breeding ground for it, amplified by foreign agents bots since Elon took over.

Re: Why Grok Fell in Love with Hitler

#9

Post by rapatel0 » Fri, Jul 11, 2025, 1:26 PM UTC

The key insight here is "too compliant to user requests"

What likely happened is that a few people decided to query groq in to generate ragebait traffic to the page/account/etc. Then critical mass happened to make it go viral. Then it confirms prior biases so the media reported it as such (and also drive clicks and revenue).

Microsoft had basically the same scandal with twitter chatbot as well a few years ago.

Sadly, ragebait is a business model.

Re: Why Grok Fell in Love with Hitler

#10

Post by PaulHoule » Fri, Jul 11, 2025, 1:26 PM UTC

https://www.youtube.com/watch?v=oDuxP2vnWNk

but wish Jay-Z would slap Ye for the squeaky autotune at the start