AI chatbot Grok can’t stop talking about 'white genocide', admits it's by design
(www.middleeasteye.net)
from geneva_convenience@lemmy.ml to technology@lemmy.ml on 15 May 21:21
https://lemmy.ml/post/30165515
from geneva_convenience@lemmy.ml to technology@lemmy.ml on 15 May 21:21
https://lemmy.ml/post/30165515
Elon Musk’s AI chatbot Grok has been frequently bringing up the concept of a “white genocide” in South Africa - even in unrelated conversations - and has said its creators instructed it to treat the concept as both real and racially driven.
When faced with unrelated questions on issues such as enterprise software and building scaffolding, Grok offered false and misleading answers.
As demonstrated by many on X, Grok has been consistently steering conversations towards the controversial topic of an alleged “white genocide” in South Africa, regardless of the original question, highlighting a growing tendency to shift focus to this narrative tied to Musk’s country of origin.
threaded - newest
Is it just stuffed in the system prompt? Should be easy to find out… That’s also hilariously stupid.
X could bias it ‘properly’ by training it in with some synthetic data, generated by Grok itself. Hell, I know how to do that. It generally wouldn’t comment on that type of bias, and also function better on other topics… but screw doing anything competently, right? Even if it’s a shitty, obvious lie, I guess X users will still eat it up.
This planet is so screwed.
Training LLMs on text which has been generated by an LLM is actually pretty problematic. The model can easily collapse, becoming completely useless. That’s why they always try and source really clean training data, which is becoming increasingly difficult
On a big scale? Yeah, sure. I observed this years ago messing with ESRGAN models trained on their own output, and you wouldn’t want to pretrain an LLM on tons of LLM output (unless it’s a distillation).
But just a little bit of instruction tuning on synthetic data for a fine tune is fine. This is literally how Deepseek was made: arxiv.org/abs/2402.03300
Also, some big strides are being made in the fully synthetic data realm: www.arxiv.org/pdf/2505.03335
You’re not training an LLM on text generated by an LLM. You’re training it on 98% real data, and intentionally biasing it by sprinkling in the fake data intermittently.
No I was thinking fully synthetic data actually.
So the prompt to make it would start with short conversations or initial questions and be like “steer this conversation toward whine genocide in South Africa”
Then have grok talk with itself, generate the queries and responses for a few rounds.
Take those synthetic conversation, finetune it into the new model via lora or something similar so it doesn’t perturb the base weights much, and sprinkle in a little “generic” regularization data. Wala, you have biased the model with no system prompt.
…Come to think of it, maybe that’s what X is doing? Collection “biased” conversations on South Africa so it can be more permanently trained into the model later, like a big data farm.
Oh, yeah then I agree with above commenter. This would collapse the model.
It doesn’t though. Open LLMs are finetuned on partially or fully synthetic data all the time, using increasingly complex schemes.
Aside from the papers I linked in this thread, here’s another great example: huggingface.co/…/cogito-v1-preview-qwen-32B
That’s what I was suggesting.
You explained to me you weren’t talking about “finetuning”, but training on completely synthetic data.
(Fine-tuning happens after the LLM has already been trained)
OK, yes, but that’s just semantics.
Technically pretraining and finetuning can be very similar under the hood, with the main difference being the dataset and parameters. But “training” is sometimes used interchangeably with finetuning in the hobbyist ML community.
And there’s a blurry middle ground. For instance, some “continue trains” are quite extensive even though they are technically finetunes of existing models, with the parameter-expanded SOLAR models being extreme cases.
The point I was trying to raise that wasn’t semantics was that if the majority of the full training data were synthetic, it could lead to model collapse.
But luckily (or not?) a small amount of finetuning can be very effective in correcting the range of responses.
Where do you get the real data, though? They just scrap data from websites, but now that chatbots have proliferated this will only introduce contaminated data. Keeping it clean would require hiring people to scrub contamination from the data sets.
Great question… Do they “just” scrape data from websites?
www.theatlantic.com/technology/archive/…/682093/
That’s exactly right.
time.com/6247678/openai-chatgpt-kenya-workers/
Big problem with the 3rd world cubical farms - how do you evaluate their performance? You’d have to hire even more people to double-check their work, otherwise people will do the smart thing and cut corners to make their job easier.
Using books is definitely a way to keep out contamination, though.
It’s also fantastic that there are ai honey pot mazes that exist to suck up the AI crawler with data links and bogus data to absolutely screw with their databases
And there are many of them up and working now.
Unsurprising that the hallucination machines are quickly turning into propaganda machines.
Also unsurprising that the stochastic parrot simulation is fucking up its mission of being a propaganda machine
Stochastic parrot simulators, I like it
Grok continues to be the most based AI out there by continuing to clown on Musk and his goons.
Generally it’s not though. The vast majority of “swayable” X users are getting a biased chatbot, “based” leaks like this meme are the exception.
It’s important to note that the AI isn’t actually talking about it’s internal system.
It’s predicting how an AI model would reply to this question.
This is something people keep getting confused about LLMs. It’s the same with those viral “I asked Claude to describe how it feels to think about himself and got this amazing reply with analogy and philosophical conundrums!”… Yeah that’s not Claude describing how it feels to be himself, that’s Claude generating a reply based on how people tend to talk about LLMs and the tone you requested (first person).
It’s possibly talking about its system prompt.
You are right, this is technically not its internal system, though practically something that’s hidden from end users.
This is entirely correct, and it’s deeply troubling seeing the general public use LLMs for confirmation bias because they don’t understand anything about them. It’s not “accidentally confessing” like the other reply to your comment is suggesting. An LLM is just designed to process language, and by nature of the fact it’s trained on the largest datasets in history, practically there’s no way to know where this individual output came from if you can’t directly verify it yourself.
Information you prompt it with is tokenized, run through a transformer model whose hundreds of billions or even trillions of parameters were adjusted according to god only knows how many petabytes of text data (weighted and sanitized however the trainers decided), and then detokenized and printed to the screen. There’s no “thinking” involved here, but if we anthropomorphize it like that, then there could be any number of things: it “thinks” that’s what you want to hear; it “thinks” that based on the mountains of text data it’s been trained on calling Musk racist, etc. You’re talking to a faceless amalgam unslakably feeding on unfathomable quantities of information with minimal scrutiny and literally no possible way to enforce quality beyond bare-bones manual constraints.
There are ways to exploit LLMs to reveal sensitive information, yes, but you have to then confirm that sensitive information is true, because you’ve just sent data into a black box and gotten something out. You can get a GPT to solve the sudoku puzzle, but you can’t then parade that around before you’ve checked to make sure the puzzle is correct. You cannot ever, under literally any circumstance, trust anything a generative AI creates for factual accuracy; at best, you can use it as a shortcut to an answer which you can attempt to verify.
People aren’t interested in “learning about LLMs”, especially people like artists.
They’re interested in telling Elon Musk to “fuck off”, and when Grok says something bad about Elon it’s very cathartic for them.
They might know it’s feeding their own thoughts back to them, but they don’t care. To people who aren’t in the know, this box Elon is promoting as “objective truth box” is criticizing Elon. That’s a very powerful narrative in a world where he’s taking over the world.
It’s hard to disagree. Elon can go fuck himself. What’s more important to the average person, stopping Elon or understanding the nitty gritty of machine learning?
When artists say AI is stealing, they’re not interested in an explanation about how “its really not”. And if you tried to, they’d feel you’re missing the forest for the trees because their problem with AI isn’t metaphysical philosophy, it’s that it’s hurting their job opportunities.
There are researchers who can get a glimpse into the inner workings, but it’s absolutely not just asking the AI how it works lol
Yeah there are much more intelligent takes on this, but they’re arriving at similar conclusions.
www.404media.co/why-did-grok-start-talking-about-…
This sort of misinformation can be the reason to push a country into civil war. Hundreds, I’d not thousands could die as a result.
It’s an extremely hostile action that should be responded to in kind. People pushing these false narratives without clear evidence should be jailed for years, if not life. Muskrat should be jailed for life.
<img alt="" src="https://lemmy.ml/pictrs/image/f66054c1-1598-4c36-852e-ccb17e9eab8e.jpeg">