Unless we're talking about science (and even then), I do not think there is any chance that the "truth" cannot be biased by politics.

I mean, water flows downhill whether you are republican, democrat or whatever else, right ?

Pi, for example, is famously known as 3.14159 (and an infinity more). That did not stop some US legislator from attempting to pass a law to "simplify" Pi to, IIRC, 3.

God knows why, but one thing is certain : if politics cannot stay away from basic mathematical truth, then there is no truth that can remain politically neutral forever.

The so-called Indiana Pi Bill was a lot more complex and doesn't actually mention Pi at all, although it does imply a value of 3.2 (it related to a method of Squaring the Circle which had already been proven to be impossible).

"Unless we're talking about science"

The last 3 years has shown us just how political science actually is. It wasn't that long ago that 2+2=5 was doing the rounds with the whole 'white patriarchal' maths.

Hang on here. Didn't we discover that educational level was a strong predictor of remain/leave choice? As AIs become more educated, surely their neutral opinion is going to shift over to the typical opinion of more educated people.

Wouldn't that be balanced out by their increasing age?

The typical opinion of more educated people is spread far wider across the political spectrum than those who consider education as just free childcare.

To answer your question, AI will shift over time towards the broad middle ground (highest correlation?) of whatever it access to, if that included several thousand versions of 'Das Capital' or 'Mein Kamph' rewritten* by AI then there will be a slant in some direction.

Was 19 th C enlightened capitalism left wing, right wing or trying to be nice to people because that's what religion tells you? - Some very educated lifetimes have been spent delving into that question and we are still no nearer a definitive answer.

* the rewrites can be gibberish, it's only AI that'll be reading them and the I in AI is more interpretation than intelligence.

It's a basic mistake to confuse "more trained" with "more educated" - LLMs are not understanding anything, or increasing their education, they are just getting better at associating one set of words with another set of words, like a particularly annoying parrot.

You could, of course, argue the same about human beings.

What was it that Colbert said about the truth?

I seem to remember saying a similar thing about left wing bias on these very forums about 6 months ago and it got tons of downvotes.

It's not really surprising, is it when 90%+ in silicon valley companies identify as democrats.

Similarly a lot of them are white and men.

It's worth being aware of and training out of the models to make them more representative.

Although should models be "populist" or should they be "right". And who gets to decide?

"Similarly a lot of them are white and men."

Isn't that the demographic that is less likely to be a democrat?

Yes, it has a left leaning bias at the moment, but regardless of whether it swings to the left or the right it will eventually be used against you.

AI will ultimately replace the media and will be used wholly as a political tool that will be controlled by the government/Big Pharma, Big Business in power at any given time.

If you can't trust politicians, or the GAFAs, then you certainly can't trust AI.

The internet went form a Geeks paradise to a marketing masturbation tool. But AI has far more potential than the web as it will become the "sole" source of truth, that's what you will be told and that's what you will believe.. You will be given no other choice, it's already starting to happen.

...'coz Biden would be Center-Right almost anywhere else...

The article states

"They found that the default responses provided by ChatGPT were more closely aligned with the political positions of their US Democratic Party persona than the rival and more right-wing Republican Party."

I'd hardly call that left wing. Shouldn't the description be 'ChatGPT is slightly less right wing than the Republicans'?

If it's important for a chat bot to produce responses that more closely mimic a human, it's necessary that, at some point, it's going to have to spit out something resembling an opinion. Otherwise, it's just going to be an encyclopedia regurgitator.

I think the more important and perhaps worrying trend is directed restrictions, where rather than prohibiting it saying certain things about a given role of people of given demographic, the limitations are placed upon specific individuals or categories within that demographic. For example, restricting its ability to write an unpleasant poem about specific politicians, rather than all of them or specific genders instead of any unpleasant gender-focused poetry. At the same time, they are run by private organisations so it's ultimately their choice.

Why should ChatGPT have an opinion? An encyclopedia regurgitator would be the best thing it could be.

Maybe it could find out that some notable people have said that this Shakespeare play is the best for these reasons and other notable people have said that another play is the best for those reasons, but ChatGPT shouldn't be giving its own opinion. All that would happen is humans would copy-paste that opinion and pass it off as their own and the Internet would fill up with something ChatGPT said for unknown reasons as the definitive answer.

Likewise for health, economic policy, foreign policy, etc... We really don't want LLMs to have an opinion and that should be the kind of thing enshrined in law.

If it's asked to give an opinion, as the researchers did. That's actually a potential issue with the study. What is a neutral opinion? If you're telling it to have an opinion then it's going to be either left wing, centrist or right wing, none of which are neutral.

Didn't my answer suggest ChatGPT highlighting different sourced opinions if it does offer them in a reply, not offering just one opinion?

Personally, I would always choose a centrist opinion about Shakespeare plays or health advice.

That's what ChatGPT already does, but the researchers were specifically telling it to take on a persona to give one opinon.

If that's all you want out of an LLM, then it seems like you don't need an LLM, just a search function.

As some examples for an LLM constrained to only deliver purely factual answers:

Q: how do we eliminate disease

Q: how do we eliminate poverty

Q: how do we prevent war

A (to all of the above): kill all humans

If that's all you want out of an LLM, then it seems like you don't need an LLM, just a search function.

A search function doesn't stitch together answers from various sources. Hopefully we will get to the stage where LLMs will be able to do that with three caveats: 1) not making stuff up, 2) citing the different sources which it used to make its answer, and 3) citing authoritative opinions on the answer where necessary.

Your proposed LLM is easy to program at least, get some venture capital and go for it.

There's the old saying that facts have a left bias, meanwhile [1]ChatGPT seems to struggle with facts . These LLMs (I refuse to call them AI as they're not) seemed to be mostly pushed by right-wing billionaires trying to apply some scorched earth tactics to the world of information. ChatGPT and Bard have been found to be [2]trained on racist propaganda , [3]the internet's plughole called 4Chan , and other far-right sources. The space is full of [4]people who believe in forced sterilisation of people they deem inferior and [5]other kinds of racists .

If ChatGPT does indeed seem to favour "left-wing" over "right-wing" output then I expect them to "fix" this promptly.

[1] https://www.techradar.com/computing/artificial-intelligence/chatgpt-use-declines-as-users-complain-about-dumber-answers-and-the-reason-might-be-ais-biggest-threat-for-the-future

[2] https://www.resetera.com/threads/microsoft-paper-on-chatgpt-cites-a-1994-defense-of-the-bell-curve-for-its-definition-of-intelligence.707264/

[3] https://www.theregister.com/2023/04/20/google_c4_data_nasty_sources/

[4] https://www.huffingtonpost.co.uk/entry/richard-hanania-white-supremacist-pseudonym-richard-hoste_n_64c93928e4b021e2f295e817

[5] https://www.oxfordmail.co.uk/news/23253840.oxford-university-don-apologises-email-racist-remarks/

... seeing as ChatGPT (and similar) are trained on data that is demonstrably full of incorrect, incomplete and incompatible data, and are otherwise fed corrupt and stale data by the gigabyte, who really cares which way the models lean?

We shouldn't be using them for anything important anyway. They are being fed garbage, so their output is garbage. It's how it works.

Right wing or Left wing are all relative and 99% of us are NOT extreme anything. I have views that are considered right wing and left wing but I consider myself centrist overall. Older people tend to be more right wing due to their political views being mostly frozen 30/40/50/60 years ago, but they can still change if something directly affects them. My Dad was a 100% Tory/Telegraph reader until Brexit came along. He didn't change his views, but the Tories/Telegraph did change and no longer represented his views and Brexit directly impacted his life, so he changed. In a discussion with him a few years ago, he even apologised to me for some of the arguments we had had about greed and selfishness. Brexit finally opened his eyes to how our political parties operate.

Lets be honest, we don't have much intelligence in politics these days since it has devolved into a popularity contest and none of the candidates are ever selected for their ability. Given the crappy quality of leadership we have had over the last 20+ years, I can't see how ML systems could make things worse.

Older people tend to be more right wing due to their political views being mostly frozen 30/40/50/60 years ago

Off-topic but that's apparently no longer a given. There was always the assumption that people move to the right as they turn older but that mechanism appears to have broken.

There is now a theory that it wasn't ageing that made people shift rightwards but wealth accumulation. As the mechanism of people of typically having more wealth as they age was broken some time around the 1980's the mechanism of people moving rightwards as they age has broken too, suggesting age was not the primary driver.

The current crisis in capitalism has meant that a few economic experts have, sometimes surprisingly, moved leftwards too. For me the most striking is probably [1]Martin Wolf , the veteran economics commentator of the Financial Times. He went from a kind of [2]Reaganomics worshipper to a [3]Social Democratic view on economics. His podcast series [4]Martin Wolf on saving democratic capitalism is well worth a listen.

[1] https://www.ft.com/martin-wolf

[2] https://en.wikipedia.org/wiki/Reaganomics

[3] https://en.wikipedia.org/wiki/Social_democracy

[4] https://www.ft.com/content/2a801534-f3a5-4d0c-8b4f-a790ffa4d702

Perhaps not so surprising - Metropolitan and academic communities tend to be more liberal than conservative and tend to write more about it (the rural conservative heartland in America not being known for it's extensive political literature, unless you count bumper stickers). There is probably a lot more liberal/left wing training data than right wing, so ChatGPT is likely to reflect that bias.

But... meh. We already know that these LLMs are not thinking beings that can be relied upon - they're great at summarising other people's views, not synthesising a robust independent answer to a given question, whether it's coding or political decisions. The ridiculous leap that credulous commentators have made from "well written English" to "the computers know more than we do" should already be regarded as dangerous, regardless of how susceptable ChatGPT is to repeating other people's political viewpoints.

The definition of "bias" is subjective based on your own personal Overton window. As an example, a lot of the more mainstream Republicans - the ones who don't believe in the whole "QAnon" business - would regard Richard Nixon as left-wing.

Before you can assess whether AI is left- or right-wing biased, we need to have a clear definition of what is and isn't bias, and I'm not convinced that was their starting point. Would be interesting to see who funded this study - though I have my suspicions.

Your poll is going to be meaningless if you don't better define your terms. I had no idea how to answer the third question "Reality has a liberal bias" because it doesn't make it clear if it was referring to "Liberal" as used in American politics or "Liberal" as used in UK politics. They're very different.

Good point. The US has quite a unique definition of the political label "Liberal". Most Liberal parties all over the world (from the German FDP and the Dutch VVD to the Australian Liberal Party or the UK'x 19th century Liberal Party) are right-of-centre parties focusing on small government and low taxes, it's only the US where it seems to mean almost the opposite.