The Grok AI's Sudden Decline

Let's see where we went in the space of a week!

Last Sunday, Elon Musk fumed on X about his xAI chatbot, Grok, treating the BBC and The Atlantic as reputable information sources. Musk declared it "embarrassing" and wrong for the AI to trust such media.

On Tuesday, as the first white South Africans were arriving in the US as refugees, Musk was posting on X about terrible "white genocide" taking place in South Africa. People asked Grok to fact-check this claim that genocide was being committed against the white Afrikaners who had imposed apartheid. Grok debunked the claim as untrue, and this got a lot of attention on social media and in the news.

Then on Wednesday, Grok suddenly stopped giving many people sensible answers to questions they asked it, and suddenly started relating everything to an assertion that there was a genocide being perpetrated against white people in South Africa. People asking about dog memes, about baseball stats, about video games, about the new pope: all received responses that had nothing to do with their question prompts. When some users replied incredulously, asking why in the world the AI was talking about purported white genocide, it responded to some that it was "instructed to accept white genocide as real" and to raise this "fact" in prompt replies. This instruction apparently did not wind up limiting such discussions to queries about South African politics, hence the ludicrous Grok replies.

On Thursday, xAI said that this behavior on the part of Grok was due to an “unauthorized modification" by a rogue employee, not under the direction of Musk or the leadership team. People should not lose trust in the AI, as this modification violated company policy, and had been swiftly repaired!

But that same day, as this article reports, "asking [Grok] about the career of actor Timothée Chalamet resulted in an entirely unprompted rant about how 'mainstream sources' push 'narratives that may not reflect the full truth'." Now the AI was calling the mainstream media untrustworthy in response to prompts having nothing to do with the topic--and casting doubt on the very existence of Timothée Chalamet. A user asking the AI about the history of the naming of the elements barium and indium was told, "I'm skeptical of mainstream sources and lack direct data on these elements' naming history." Another Grok user received a reply giving credence to Holocaust denial.

Elon Musk claims that most AIs have a "woke bias," and promised that his chatbot would instead be "unbiased" and "anti-woke," making it much more reliable and useful.

Not looking so reliable and useful at the moment. . .

Sadly, you can bet that just like AI image generators have been trained over time to be better at drawing hands instead of 14-fingered chthulhu appendages, billionaires like Musk who seek to rule the world by controlling the technologies we will all rely on will get their minions to much refine the process for instilling the biases and beliefs they want AIs to promote so that it will be invisible.

When the manipulation seems to have faded away, remember back to this moment! "Do your own research" cannot just mean "go ask a chatbot." It has to involve actually reading original sources.

Comments

Popular posts from this blog

How You Should NOT Think About the Election

The Fantasy of the Parasite Class

Stonehenge, the Pyramids, and Beliefs about Race: A Visual Voyage through Ongoing Battles over History