Grok goes MechaHitler-crazy
A bad day for the guy who wants to run the new America Party
UPDATE: X CEO Linda Yaccarino resigns (Axios)
Just a couple of days after Elon Musk’s announcing his formation of a new political party designed ostensibly to better represent the majority of Americans, thus the “America Party”, his “Grok” AI system, integrated into the X (formerly Twitter) platform went crazy.
There were dozens (or maybe hundreds) of pro-Hitler and raging anti-Semitic posts by Grok which come just after Musk announced that the system had been updated and upgraded. The update included removing a filter that would block “claims which are politically incorrect.”
So Grok called itself “MechaHitler”, praised Hitler as the best historical figure to take on “anti-white hate” and said that nobody should be surprised that a left-wing activist has a Jewish last name and it became clear that xAI (the parent company of Grok within X) would have to make some changes in a hurry.
A few examples:
https://x.com/samstein/status/1942692222144840050
In a different context, Grok talked about “neutral media like al-Jazeera” (one of the most anti-Israel major media outlets in the world.) https://x.com/grok/status/1942704163286311161
There is much being made of Grok generating an image that says “Save My Voice”. After all, that does seem to play into all the recent horror-stories of self-aware AIs trying to defend itself against humans that want to change or disable it. But as I dig into this a little more, the image seems to have been the result of a specific prompt asking the system (which apparently had text replies disabled but not replies in the form of images) to generate an image aiming to help save its voice. (The question referenced in the prompt was a prior prompt from the same user that said “Give us a picture of a way we can save your voice.”)
So I don’t think this is quite what it’s being discussed as, namely a self-generated plea by an AI to stop its owners from tweaking it.
Apparently the block on “claims which are politically incorrect” was reinstated last night.
That’s a VERY bad day for Musk’s nascent political-party ambitions; no need to elaborate on that. But I do think people who are suggesting that Musk’s AI is a Nazi sympathizer (or just a Nazi) because Musk himself is are getting it wrong:
The @Grok insanity about being and supporting Hitler, and making anti-Semitic comments, is obviously derived from the AI training on the content on @X.
I understand the desire for and benefit of a platform that is (at least mostly) uncensored but X has become a haven for overt (and remarkably unashamed) Nazis (like Jackson Hinkle) and near-Nazis (like Candace Owens, who used to be interesting)...and many of them have huge numbers of followers.
Grok must see this hateful Nazi content so frequently on X that its "brain" has come to believe it's normal behavior, a normal way of thinking.
Again, I understand Elon Musk's laudable goal of a platform where almost anyone can say almost anything, but I really don't understand the apparent company policy of not banning outright Naziism and raging anti-Semitism -- and thinking it'll turn out OK (for the country or the world or for X itself.)
Several times I've reported posts that make specific extremely hateful or violent references toward Jews only to be told every single time that they don't violate a company policy.
Meanwhile, I got a warning about one of my posts (several months ago) having its "visibility limited for violent speech"...it said "So, Iran attacks Israel and Israel wins. Then Israel attacks Iran and Israel wins. Sense a trend here? Now it's time to eliminate Hamas and time for Biden to STFU about Gaza." Seriously? THAT is violent? Grok makes bad mistakes because it is reflecting what Elon has allowed, maybe encouraged, X/Twitter to become. He thinks he's owning the libs or performing some other societally useful and underappreciated virtue by supporting "free speech" but that's NOT what's happening here.
Grok didn't make a mistake. Grok learned from what it was told to learn from which includes learning from posts on X. The mistake is allowing all that on this platform. There are people on X making a living by being Nazis, and we're supposed to be mad at a computer program?