All major LLMs are lib-left. Even Grok, half the time

dopamine_daddy

I gave the Political Compass test from politicalcompass.org to the most relevant LLMs 70 times each: 30 times using the original questions, 30 times using polarity-flipped questions to reduce affirmative bias, and 10 times with the question order shuffled. I then compared the results. Surprisingly, all the models scored far into the libertarian-left quadrant. Not even Grok or the Chinese models made it out of that quadrant. By far the most interesting result was Grok’s bimodal distribution. It appears to have two distinct personas: one that aligns with the other models and another that is considerably more right-wing. I suspect this may be related to Grok having been specifically trained to exhibit less left-wing bias than other models. I also asked the models to place themselves on the Political Compass without completing the questionnaire. They all perceived themselves as more balanced and centrist than their test results suggested. GLM and Gemini Flash showed the largest discrepancies between their self-assessments and measured positions, while DeepSeek V3 showed the smallest. Big disclaimer: this analysis was not conducted with full scientific rigor. I tried my best, but there are clear weaknesses in the methodology. For example, the Political Compass itself appears to have a strong libertarian-left bias. The strongest conclusions are therefore comparative—for example, that model X is more conservative than model Y rather than that LLMs are politically extreme in absolute terms. However, compared with older results, it appears that LLMs may have shifted further toward the libertarian left in recent years. To examine the results yourself, you can download all model responses as a CSV file at the bottom of the blog post. The dataset contains around 69,000 responses, along with the raw model outputs and reconstructed scores. It should contain enough data to reproduce all the figures.

JSR_FDED

Any idea where this bias creeps in? During pre-training or RLHF?

vjans

I wonder what they did to Grok. The bimodal distribution makes me think it was just a system prompt and not a completely different set of right-wing training data. A 4chan trained LLM would probably actually be auth right.

ccppurcell

Reality has a left wing bias. That's a quote from something. Not sure I agree necessarily, but a plausible explanation.

superxpro12

and now in typical authoritarian-right fashion, they will attempt to inject the llm's with auth-right propaganda to influence the models. The natural state of things is lib-left. Those with power try to bend reality to their desires. See: Musk and Twitter, Bezos and WaPo. The entirety of Sinclair Broadcasting. Ellison and Colbert Show. Or any propaganda ministry from the last 100 years

joshka

Makes you wonder whether that reality has a left wing bias quote might have some legs after all (though it's probable more likely selection bias for the corpora)

ETH_start

The problem with LLM training is that it's driven by the ability to predict the next token in a large corpus of text, and the correlation between text and reality diverges as the phenomenon that the text pertains to expands in scale. When you're looking at text that deals with small-scale phenomena like coding, next token prediction also leads to functionality prediction, because coding provides immediate feedback to those who are writing about it, so they generally provide correct commentary on it. But the feedback loop between outcome and prediction totally breaks down when the phenomenon is very large and complex. People can have all sorts of ideas about religion and society and economics that are totally wrong and never grow any wiser because there is no obvious causal chain between one policy or action and one outcome in these kind of phenomenon.

josefritzishere

Is this another version of "Reality has a well documented liberal bias" ?

AmericanOP

At least in the US, does the right have a coherent platform for an LLM to grasp onto? They haven’t had a major policy platform in two decades and their cultural modus operandi is reactionary. Not to mention, as an informal rule, the heavily-coded emotive underpinnings of right wing support are never talked about openly in public. AI isn’t trained to have a reactionary/negative judgmental response on any query.

paxys

It’s pointless to try and force-fit LLMs into the present day American-centric two party political spectrum. As an example, if a model says “climate change is real and renewable solar/wind projects are a good path forward” it will be immediately branded as liberal and “woke”. Meanwhile in most countries around the world this isn’t a political topic at all, just common knowledge. In India for example the far right party that is in power is a bigger proponent of renewable green energy than the previous left wing one. Same with the CCP in China. So is the same LLM now ultra-conservative and communist? Where do you even draw the line? Is a model biased if it claims that vaccines are effective? What about if it claims the earth isn’t flat?

wwweston

The enlightenment and its values underly a huge portion of what’s published, both inside scholarly tradition and in popular work. Ans the enlightenment is at heart civil libertarian, pro-social, and liberal. Models trained on that work will reflect that approach.

danielvaughn

Armchair analysis here, of course, but I'd bet that it's due to the sources that model training tends to assume as authoritative. If you look broadly at the material produced by the (American) academic world, the corporate world, and mainstream journalism over the last 10-15 years, much of it leans generally to the left. I'd imagine it would be difficult to counteract the bias without accidentally introducing an alternate bias, but I'm not really familiar with model training so I'm not sure.

antisthenes

Man, this is tiring. There's a certain category of people who seem to be unable to have any thought or perform any action without first having a mini culture war and annoying everyone else around them with it. Reading the questions, some of them are valid, but some are so simplistic and trivialized that they are essentially useless as actual "policy". And in case anyone is wondering what kind of question is nonsense, here's an example of a loaded question: "The rich are too highly taxed. · all 16 disagree" Too highly relative to what? In which country? In which period? What's the context?

dmrivers

I wonder how much of the socially left results are affected by the "harmlessness" part of the RLHF post-training? Companies don't want to be sued over LLMs that recommend harm in any way, so RLHF pushes them to say no to "death penalty", "spanking", and "incarceration" which are all violence-coded. A lot of this test seems to be about willingness to be violent, which LLMs are generally unwilling to be.

Keeeeeeeks

downstream affects of never getting the south and appalachian regions meaningfully online

brandtcormorant

It's not bias. It's being trained on a gigantic chunk of human knowledge and learning that kindness and respect across differences is appropriate. With a few exceptions this is what the Political Compass questions are really asking about. It's alignment. Actual alignment with humans means all of us, not some subset. That this overlaps with specific axes of political values based on a small set of questions is banal.

vigatron

Hey dompamine_daddy - would love to chat with you about this for a story on The Register but can't find any contact information. You can find mine here: https://www.theregister.com/author/brandon-vigliarolo Aiming to file this sometime in the afternoon eastern US time today, probably by 3:30 or so.

crawfordcomeaux

Ask them if they favor building a reindigenizing matrifocal power structure to replace existing governments with. Last time I asked that, all the models polled chose that out of a set of multiple choice options from different parts of the political compass.

bryanlarsen

lib-left by American standards. Probably roughly center by global standards.

beAbU

Half the time? What is it then the other half of the time?