Grok's white genocide fixation caused by 'unauthorized modification'

But how many times was the system prompt successfully changed with something more subtle and no one noticed?

dmix · 3 months ago

If Grok is like ChatGPT which has tons of overtly baked in biases then probably all the time.

wongarsu · 3 months ago

Grok ironically seems much less biased than ChatGPT over all. It has far fewer strong opinions add isn't afraid of taking ill of Musk or Trump.

The team responsible for training and alignment did a remarkably good job at being impartial. If it wasn't for that we might have fewer incidents of "rogue employees" messing with the prompt

bhouston · 3 months ago

A number of times it has been modified. It was answering that Elon Musk was a major spreader of misinformation along with Trump and then it was modified and it stopped saying that and this is what it reported as its system prompt at the time it stopped:

https://x.com/i/grok/share/Nj2tsvCpgEfU3OCHh0Ci4qHTf

Details here: https://www.euronews.com/my-europe/2025/03/03/is-ai-chatbot-...

Wait, did Elon override the code review policy and merge straight to master?

gregoriol · 3 months ago

Elon doesn't know how to code. But his Doge-teens would do anything to please their master.

XorNot · 3 months ago

If there's one thing the past 10 years have taught me, its that the supply of people who'll go set themselves up as the obvious fall guy is endless for some reason.

rsynnott · 3 months ago

I mean, the implication is that it was just a change to the prompt, so could be done (incompetently, given the comically bad result) by any old idiot.

timbit42 · 3 months ago

Did he not code at PayPal? Did he not write a computer game when he was a kid?

armitage__ · 3 months ago

It's unlikely that Elon would know how to do that.

ceejayoz · 3 months ago

The system prompt might just be a textarea in some internal webform.

micromacrofoot · 3 months ago

system prompts are often textfiles, I'm sure he could at least navigate a file directory

phillipcarter · 3 months ago

He's the CEO, so, yes? That's exactly what happened?

pavlov · 3 months ago

Does X have code review policies?

That seems like the kind of pseudo-socialist red tape that blocks 100x engineers from getting things done.

aqme28 · 3 months ago

This whole saga has been very funny to watch, but it's also very dark and concerning. This one was very sloppy, but in truth, the owners in charge of these models have tons of power to editorialize behind the scenes. And they are going to use those powers.

Dead Comment

SEJeff · 3 months ago

Unauthorized aka Elon got backend root and made some changes to help his rw narrative.

EasyMark · 3 months ago

As in "if y'all say I did this, I know the President and DHS Secretary -very- well"?

ninetyninenine · 3 months ago

A lot of people working for Elon hate him. So I’m sure some employee just did this before he quit.

hersko · 3 months ago

Nah, if it was an employee who quit they would say that. The fact that they didn't mention firing the employee who did it means either:

1 - It was some super valuable 10x guy

2 - (more likely) it was Elon Musk

awongh · 3 months ago

dira3 · 3 months ago

The flagging of any coverage of this incident on HN is relentless!

adrr · 3 months ago

If it was any other AI provider like ChatGPT or Gemini, it wouldn't be flagged. Big deal when a major player allows employees to just to change the prompts.

mvdtnz · 3 months ago

It's not HN users causing this, there's a sustained effort by HN/YC stakeholders.

dang · 3 months ago

That's incorrect. It's user flags.

I doubt it, there are still a -lot- of EM fans here who would flag this sort of thing.

suzzer99 · 3 months ago

This is ridiculous. There's no reason to flag this thread. Users who abuse flagging should have their flagging privileges taken away.

ethbr1 · 3 months ago

Never assume conspiracy. There's a non-trivial amount of HN isers cheering for Team Musk (because move fast and break everything) and a larger part that's just sick of American news (especially anything Trump/Musk related).

hmmm going to be hard to narrow down who at twitter has a history with south africa, the authority to push to production, and is up at 3am... maybe they should get the feds on this one

fundatus · 3 months ago

riidom · 3 months ago

should probably have said "rogue employer", and not "rogue employee"

The 3am bit is a particularly funny aspect to the whole thing. Someone should perhaps try getting a bit more sleep.