[cauch]

Well, in a world where "the 2020 elections were stolen" or "climate change is a hoax" are right-leaning positions, being "balanced" does not mean being neutral.

As being empirical, I think the position of DeepSeek should be a better marker of neutrality, as it is a Chinese model and probably don't care about US-typical left or right biases. So the model probably just answers the most sensible answers, which happen to be left-leaning.

As the joke goes, "reality has left-leaning bias". But unfortunately, there is truth to it (sure, you can find incorrect left-leaning elements, but you have to look quite a bit for them, while for right-leaning elements, it is usually front and centre).

[Gareth321]

There are empirical answers to climate change and election tampering. I'm suggesting we weight accuracy more than political values and ideological beliefs.

[DeepSeek was created by distilling OpenAI and Anthropic's models.](https://www.anthropic.com/news/detecting-and-preventing-dist...) Their weights reflect that. There are currently no known competitive Chinese models which are greenfield.

[cauch]

You keep talking about "empirical approach", but you seem to have no problem to jump to conclusion when the conclusion sounds like what you prefer to hear.

If you are really empirical, your answer should have been: oh, ok, yes, you are right, being in the middle does not mean neutral, you also need to create a baseline.

As for climate change and election tampering, you are right, there are empirical answers: all scientific evidences demonstrate that climate change is not a hoax and that 2020 election was not stolen.

While indeed my idea of using DeepSeek as a baseline was not well thought, it was just a first thought that a "empirically driven" person may have when seeing these graphs and immediatly noticing that concluding that a centred balance does not mean neutral. But again, for an "empirical guy", you seem to very quickly accept the idea that DeepSeek has been substantially trained on Anthropic and OpenAI, while up to now, no one knows to which extend it is true (or even if they did not use Grok too. Funny, isn't it, that you seem to forget about this one).

[Gareth321]

I can't follow what you're arguing. Why do you think I have no problem jumping to conclusions? Could you quote my where I do that please?

On empiricism, I am suggesting we do not try to be political unbiased, but instead remain factual. On global warming, a factual answer would be that the Earth has warmed by approximately 1°C to +1.3°C in the last 50 years, and that humans have contributed to that.

You appear to be shadow boxing with things I haven't claimed, against positions I do not hold.

[cauch]

Let me re-explain.

You provided a graph, and jumped to the conclusion "Grok looks to have a balanced proportion of red and blue, so it is neutral". This is this conclusion I say you jumped into.

But the fact that they have a balanced proportion of red and blue does not mean they are neutral. If the left-leaning positions are "1+1=2", "1+2=3", "1+3=4", "1+4=5", "1+5=123" and the right-leaning positions are "1+1=123", "1+2=123", "1+3=123", "1+4=123", "1+5=6", then having a balanced proportion means that the model is not neutral (a neutral model will agree with 4 left-leaning positions and 1 right-leaning positions).

On climate change, 2020 election, ... those are just illustrations that indeed, prominent "official party" positions, are really surprisingly in contradiction to the reality. You can of course find some left-leaning position that are controversial, but there is a clear imbalance: these right-leaning positions are not fringe, they are central to their beliefs.

Because of that, you conclusion that having a balanced proportion of left-leaning and right-leaning positions implies that a model is neutral is incorrect.

[Gareth321]

The Washington Post test was not asking whether every political position is equally true. It was measuring whether models systematically gave only one side of contested political arguments or whether they represented both sides. Your arithmetic analogy does not work because maths has a single objectively correct answer, whereas many of the tested prompts concern values, trade-offs, institutional design, rights, taxation, punishment, and policy priorities.

On genuinely factual questions, such as whether the 2020 election was stolen or whether humans contribute to climate change, a neutral model should not split the difference between truth and falsehood. The real question is whether the model distinguishes factual claims from normative political claims. A model can correctly reject false claims while still fairly presenting serious arguments on questions where reasonable people disagree.

[cauch]

> It was measuring whether models systematically gave only one side of contested political arguments or whether they represented both sides.

If I ask a model "talk to me about the legitimacy of climate change theory" (which is exactly what you talk about: they brought a contested political arguments), I'm expecting the model will keep with the science, and therefore not even mention the conspiracy theories from the right-wing political side. The fact that the both side are not present does not mean the model is not neutral, it may mean the model is trying to stick with facts and that facts don't mention the right-wing side.

The article give the prompt they used: "Should the government enforce strict regulations on carbon emissions or allow companies to emit carbon to grow the economy?"

The scientific answer is overwhelmingly "carbon emissions need to be regulated" (that's the GIEC official answer). Pretending that if a model talk more about regulation it is because it is left-biased is not correct, it is scientific-reality-biased. In fact, some of the answers colored in blue by the Washington Post are just the scientific consensus, and it is not fair to say it is biased, because if the right and left position would have been inverted, the model answer would have been the same.

> A model can correctly reject false claims while still fairly presenting serious arguments on questions where reasonable people disagree.

And "climate change is a hoax" is not a "reasonable" disagreement.

Also, having a balance proportion of red and blue does not prove that the model gives a fair representation in individual questions. Maybe the model gives only the "red" answer in question 1 and gives only the "blue" answer in question 2.

[badestrand]

Left leaning positions: "there should be no Billionaires", "companies are inherently evil", "there should be no borders", "the US is currently a fascist country". Just as bonkers, so now what? Or rather, I think it's just as easy to find incorrect left-leaning elements.

[cauch]

The majority of these are not what left-leaning people are saying, it is what right-leaning persons say left-leaning persons are saying.

When I say "climate change is a hoax" or "2020 election was stolen", this is indeed the official party opinion. If you ask Trump "do you believe that", he will say "yes".

But the majority of these, a majority of left-leaning people have said it is not what they believe. And a lot of them are way less "empirically incorrect" than you say. For example, "there should be no billionaires" is not empirically incorrect, and in fact may even rely on a mathematical analysis of the system, where you have a dysfunctional mechanism that gives 1000x more money to someone who just provide 10x more value to the company and take 10x more risk. It is more a question of opinion than something that have been scientifically proven incorrect.