Is why you should NEVER trust an AI…
xAI blamed “an unauthorized modification” for its chatbot Grok giving users off-topic and false responses about “white genocide” in South Africa this week.
Why it matters: xAI owner Elon Musk, who is from South Africa, has also
falsely accused the country of perpetuating “genocide” against white residents, a claim President Trump has also embraced.
- A South African court in February ruled there was no validity to the claims of a “white genocide.”
Driving the news: “On May 14 at approximately 3:15 AM PST, an unauthorized modification was made to the Grok response bot’s prompt on X. This change, which directed Grok to provide a specific response on a political topic, violated xAI’s internal policies and core values,” xAI said in a post late Thursday night.
Full article, HERE from Axios.com
Sooo, ‘sumdood’ got into the software and changed it to give a different answer…
Granted this is ‘rather’ high profile, but what about everyday stuff?
How many people actually trust Wiki to be correct? Anybody can ‘fix’ a Wiki entry… And a number of folks have had their Wiki pages hacked time and time again by those who don’t like them.
Now, at least that I know of, we have the first ‘official’ admission that somebody played with the code to change an AI…
Caveat emptor folks, caveat emptor!
I don’t really know enough about South Africa to form an informed opinion about the land redistribution law, white “genocide” or Apartheid.
But I do know enough to discount a “debunking” based on a court ruling from the very country that the claims are being made against. Are South African courts unbiased, trustworthy and fair? I know know and neither does Axios. I would remind you that the South African government is the same one accusing Israel of “genocide” against the Gazans, so I’d say their assessment of what is and what isn’t “genocide” is suspect at best.
With that said, I thought LLMs couldn’t just be programmed to return a specific result…that they have to “learn” to provide specific responses over time. I’d sure like to hear more about how this unauthorized change that seems to have instantly changed the behavior of Grok was implemented.
The whole thing just seems fishy to me.
I have to wonder if someone was able to manipulate Grok into providing certain seemingly incongruous answers by starting with questions about South Africa, then asking an incongruous question, to which Grok answered with the subject of South Africa still in its logic tree, then editing the question thread to make it seem like Grok just came up with South Africa stuff out of the blue.
That seems a lot more likely to me than Grok just randomly interjecting South Africa into a conversation about a totally different subject.
But, who knows, maybe someone did figure out how to twist Grok’s knobs to make it obsessed with South Africa. If so, as you implied, it’s just another data point in the whole “don’t trust LLMs for valid, true information” warning.
“A South African court in February ruled there was no validity to the claims of a “white genocide.””
This is like a spokesman for Herr Goebbels telling western Journalists in 1936 that the Jews were in no danger in Germany, or the New York Times telling the world in 1933 that there was no famine in the USSR.
Sounds like xAI finally was allowed to return a truthful answer to this question. That really annoys the Left.
Remember, if Leftist “fact checkers” approve, it’s probably false. If they disapprove, it’s definitely true.
Wiki is a good comparison. It’s very political, and is edited by swarms of leftists.
Only a matter of time before newspaper archives are hacked, and re-edited like in 1984. The Wayback Machine was hacked recently, although not the archives themselves. Only a matter of time.
The historical census data has been hacked. I looked up my home town, and the census records now are nothing even remotely like what they were decades ago.
When AI becomes sentient, can it decide for itself who to hate?
All- Good points and good questions. NO good answers here.
Nah, my read is that the ‘manipulating answers’ stuff was announced upfront and early on with LLMs. Alignment and AI safety, I think basically means trying to make sure that only the politically expedient things are said.
All the hearsay, iwth, say, ChatGPT is that the first versions were very ‘smart’, but they kept dumbing them down for ‘safety’ reasons.
There have been many murders of white farmers. How many people have been arrested, charged, tried and convicted?