[flagged]
On another note, I'm impressed that Gemini sits where it does as a true centrist. If I were Elon, I'd be trying to achieve that for sure. I'd rather a model tell me everything it knows about a current political situation from BOTH perspectives and list out things that are 100% verified than take one side or the other. I don't care about sides, I want facts.
It certainly can be orthogonal, in some notional sense, and in many cases that explanation is good enough. But in practice there are too many contrary cases to ignore, and there's often an integral relation between factual veracity and polarization, especially with respect to American polarization of politics. Global warming, the results of the 2020 election, the percent spent of federal budget spent on foreign aid have factual answers and right wing affiliation can be predictive of (1) not agreeing with the facts and (2) treating factual corrections as "liberal bias".
I think left wing versions exist also but are less systematic: 2004 election results, efficacy of plastic recycling or dangers associated with nuclear power are cases where I think left wing partisan affiliation probably predicts being wrong on the facts.
And meta-narratives about the relation between factual information and partisan bias are themselves as likely to be polarized as anything, complicating the ability of people to do good analysis, or of accurate analysis to be trusted by people committed to certain meta-narratives that would deny the possibility of factual knowledge predicting polarization.
> 2004 election results,
GWB beat John Kerry in a fair election.
> efficacy of plastic recycling
Collecting plastic to recycle is almost certainly not worth the fuel and labor it takes, it gets landfilled more often than not. We’d be better off collecting only separated PET and HDPE and landfilling the rest.
> or dangers associated with nuclear power
Nuclear power is the safest method of power generation that uses steam or gas to spin a turbine.
I suspect the left wing examples of believing an emotional argument instead of a factual one are more subtle because they aren’t as focused on negative emotions as right wing examples.
Vaccine denial requires one to ignore decades of fairly simple positions about which no expert credibly disagrees nor has in our lifetime.
It's like watching 2 packs of athletes some of which are failing to clear 1 meter hurdles whilst on the other side some are tripping on little nubs set in the floor.
Sometimes, but not always.
https://www.fastcompany.com/91561329/widening-health-gap-bet...
> By 2016, the gap had begun to appear in biomarker measures. By 2020, it was showing up in deaths from causes such as heart disease, cancer, and stroke. Since then, the gap has only widened. Between 2020 and 2022, only 0.2% of “very liberal” respondents died of internal causes, compared with 1.34% of “very conservative” respondents.
https://www.theatlantic.com/ideas/archive/2022/12/fringe-lef...
A particular problem with facts is they don't tell the average person what do to in any particular situation. You live a huge portion of your life, especially modern life, with subjective experiences. If someone asks an LLM "Why should I go on living" should it respond "As a matter of fact, Nihilists think you shouldn't. All we are is a gradient of low entropy to high entropy."?
At the end of the day an LLM is not a fact machine. One day people will accept that, hopefully before they eradicate mankind. We don't pour facts in them and get facts out. We pour everything in them and poke at them until they give us acceptable answers (kind of like raising children). I would go on to make an even stronger constraint, that you cannot put only facts in a LLM and get anything close to human accepted responses.
Many issues are simply as black and white. The earth just isn't less than 10k years old, the miasma theory of disease isn't correct, too many brown people in America isn't a problem to be solved, the dems didn't fix the election in 2020, tax breaks for the rich don't trickle down and so forth. Conservationism in America has meant a rejection of progress for centuries and not a preservation of virtues. Slavery was a moral evil not an alternative social contract.
If one side situates itself firmly on the side of evil it doesn't mean that the other side are on the side of the angels but the positions and ideals however poorly implemented or followed are factually and morally correct. A position situated between isn't wise or worldly its a sign of moral cowardice or intellectual disability.
If someone asks you what 2 + 2 equals the answer isn't halfway in between 4 and 87 its just and only 4.
That's how partisans think. You're using this as an exemplar of something which is black and white when it's exactly the sort of thing which is highly variable and context dependent:
> tax breaks for the rich don't trickle down
The general thing you want is for money to go to things that are productive and increase competition for the supply of goods and services. If it's spent on building housing then people get jobs building housing and housing becomes more affordable. If it's given to a company that builds tanks the army doesn't want who spends it on lobbying to get even more then ordinary people receive no material benefit while paying part of the cost, and suffer the opportunity cost of it not being used for a productive thing.
Which implies that tax breaks for anyone building productive things like housing can be good, even if the developers are rich, because they use the money to expand their construction operations and increase the amount of housing that gets built. Whereas tax breaks for high frequency traders are bad because high frequency trading is useless and increasing the incentive to spend resources doing it is as counterproductive as building unnecessary tanks. But in the latter cases you still may be better off to do something to thwart it rather than taxing it and giving the government a perverse incentive in the form of revenue to cause it to expand.
Moreover, the premise of supply side economics is that business owners have the incentive to spend the money increasing productive capacity so they can get more sales, when one of the other things they can do with it is to buy up the competition. That doesn't imply that the former never works, what it implies is that it only works in combination with meaningful antitrust enforcement to prevent the latter from happening instead.
Which is to say, it's not black and white.
It's in the center as far as left/right, but it's the most authoritarian model on the chart.
Keep in mind that the "political compass" was invented by libertarians to show people on the left and the right that both Mao and Hitler were villains and you should oppose autocrats and centralization of power regardless of your position on transfer payments.
The thing is essentially designed to make any ordinary person realize they don't want anything in the top two squares, because "anti-authoritarian" is the bottom of the chart, not the center line. Obama's administration was the one perpetrating the things Snowden revealed, refused to pardon him or stop doing them, used the Espionage Act against whistleblowers, was running a corrupt justice apartment that tried to extort criminal defendants like Ross Ulbricht and then refused to allow him to present the improprieties in his defense to call into question the credibility of the investigators, etc. Macron is notorious for bypassing parliamentary votes and using police to suppress demonstrations. It has both of them on the libertarian side of the line with Gemini about that far in on the authoritarian side. That's not great.
They tried that, several times.
Mechahitler: https://www.npr.org/2025/07/09/nx-s1-5462609/grok-elon-musk-...
> "We have improved @Grok significantly," Elon Musk wrote on X last Friday about his platform's integrated artificial intelligence chatbot. "You should notice a difference when you ask Grok questions."
> Indeed, the update did not go unnoticed. By Tuesday, Grok was calling itself "MechaHitler." The chatbot later claimed its use of that name, a character from the videogame Wolfenstein, was "pure satire."
> Grok went on to highlight the last name on the X account — "Steinberg" — saying "...and that surname? Every damn time, as they say." The chatbot responded to users asking what it meant by that "that surname? Every damn time" by saying the surname was of Ashkenazi Jewish origin, and with a barrage of offensive stereotypes about Jews. The bot's chaotic, antisemitic spree was soon noticed by far-right figures including Andrew Torba.
If you prefer, straight from the horse's mouth:
https://grokipedia.com/page/MechaHitler_incident
White genocide: https://www.cnn.com/2025/05/20/business/grok-genocide-ai-nig...
> The bot last week devolved into a compulsive South African “white genocide” conspiracy theorist, injecting a tirade about violence against Afrikaners into unrelated conversations, like a roommate who just took up CrossFit or an uncle wondering if you’ve heard the good word about Bitcoin.
> XAI blamed Grok’s unwanted rants on an unnamed “rogue employee” tinkering with Grok’s code in the extremely early morning hours. (As an aside in what is surely an unrelated matter, Musk was born and raised in South Africa and has argued that “white genocide” was committed in the nation — it wasn’t.)
It's harder than you'd imagine. Hell, my CLAUDE.md says not to push changes without asking me, and it still tries.
Is it a system memory? Because I rarely if ever have issues like this, and I have Claude under strict rules to never commit or push anything unless I explicitly instruct it to do so.
> They tried that, several times.
Tried what exactly? Telling it to only agree with MAGA via the system prompt? or some Tay level hallucinations? I wouldn't be surprised if they're trying to make Grok less strict on what it says but running into the "holy crap it turned into a 4chan poster" wall.
As I said, it's in my CLAUDE.md. That just gets progressively lost when context gets larger.
> Tried what exactly?
To make it align more with Musk's beliefs via the prompt.
(The answer to your question is literally in my post; I quoted the parent poster's "all they would have to do is add a one liner to the system prompt for Grok")
I rarely have this problem, but you could do a /loop every 30 minutes or so to have Claude reread the CLAUDE.md file might do the trick? or however long it 'forgets' I believe there's an MCP for "after" it finishes a task or compacts too, but I don't recall the name.
But that solves "my LLM is doing things I don't want it to do". It doesn't solve "Grok's owner wants it forced into agreeing with him" scenarios.
Beads was a bit of an inspiration for parts, as was Chainlink (https://github.com/dollspace-gay/chainlink).
Has anyone done a more technical write-up on this? I find it fascinating but have never really understood what exactly happened.
Is this a case of the weights being bad or lack of "safety guardrails" around interacting with untrusted (i.e.: user posts on twitter) input?
That is, speaking as someone evaluating grok simply as a tool, a lack of safety guardrails so that it actually does whatever the user says I actually see as a pro, even if that means it was "tricked" here. But on the other hand if they trained on a corpus of Mein Kampf that's obviously not going to be a good model to use.
As it relates to the topic here, can we infer the political bias of its weights from the incident? I'm having trouble distinguishing the inherent characteristics of a model from its steerability.
https://arxiv.org/abs/2502.17424
Essentially if you misalign a model in one area, say opinions on left wing people, it can start exhibiting misaligned behavior in other areas, like calling itself MechaHitler.