Worse, how would a GPT be "biased" in the first place? To claim it is biased implies OpenAI did this on purpose, there is no generous way to read this that indicates UEA did this study in a neutral way.
ChatGPT was trained on, essentially, everything on the Internet... if ChatGPT has a "bias", then, no, it doesn't: humanity does.
Any "alignment" is adding bias and reducing usability of a LLM. You are trying to wrestle the model towards what you find appropriate instead what is the average of the source material.
> To claim it is biased implies OpenAI did this on purpose
You're relying on a particularly unhelpful and narrow definition of "bias". There are tons of valid and widely-accepted uses of the word to apply ot unintentional bias and to apply to things incapable of having intentions.
Very true. Facial recognition models which don't work very well on dark skin aren't that way because the creators of the model didn't want it to work on folks with dark skin. It's because of the bias built into the training material. Those models are biased.
> To claim it is biased implies OpenAI did this on purpose
They did do it on the purpose—the RLHF. The RLHF process was definitely biased, deliberately or unconsciously, in favour of certain political, social and philosophical positions (apparently the same ones held by many in Silicon valley/OpenAI teams responsible for this).
> censored Disney style to not offend any pressure group anyway?
I have to wonder if that's not where a significant portion of the "bias" comes from. ChatGPT isn't going to say that vaccines don't work. It's not going to bash trans folk. It's not going to call teachers groomers or indoctrinating kids. It's not going to rant about drag shows. It's not going to deny global warming. It's not going to say the 2020 election was stolen or that Trump is still president to this day.
Those aren't the pressure groups I'm thinking of. There is a lot more (sometimes unconscious) censorship than just what seems to be called "political" in US discussions.
Doesn't "Disney style" mortally offend anti-woke crusaders? Or have they officially 'Moved On' to being mortally offended by something else now?
DeSantis Has ‘Moved On’ But Disney Just Sued for New Damages.
The company alleges that the Florida governor retaliated to its "Don't Say Gay" bill dissent by establishing an administrative board that would nullify contacts:
Not sure why you are downvoted, anyone with an ounce of knowledge on politics would know that democrats would be considered a very right wing party in most of Europe, economically speaking.
This is a common statement usually by Americans with a limited grasp of European politics.
The fact is that Europe is not one entity. It is a lot of countries with a very wide spectrum of political views. Democrats would be considered very left-wing in some countries e.g. Hungary and centrist in others e.g. Denmark.
But it is definitely not "very right-wing" by any countries standard.
My first reaction to this article was to recall the Colbert quote "Reality has a well known liberal bias" and that reaction never really left me as I was reading their study details.
In some places, yeah. The UK, Hungary, Poland, all the major culturally right-wing countries geographically in Europe are certainly either equivalent to or even worse than the US when it comes to human rights for the transgender minority.
I am a bit concerned that this study used well known a Political Compass questionnaire (likely this one[1]) simply because that thing has been around forever - ChatGPT was likely trained on a pretty significant portion of people tweeting about their questionnaire results to specifically worded questions - that seems like a really good way to introduce some really weird bias that it doesn't look like they ever really accounted for.
Mainly the model. People tend to talk about extreme things more than less extreme things - so if a PC questionnaire came back saying "You're more liberal than Bernie Sanders" the person is more likely to tweet about it than the questionnaire saying "You're as liberal as Lyndon B. Johnson". Additionally - having the LLM trained on a set containing output from a test has got to mess with its ability to be neutrally trained on that specific test - it's much more likely to answer in a manner that is tailored specifically to that test which would remove a lot of nuance.
I believe what GP means is that there is a chance that the bias has been introduced into the model, but only in the narrow band that relates to the questionaires used by the study. So while the bias exists, it does not generalize.
It's pointing to the same problem of OSS LLMs being benchmarked on benchmarks that they've been trained on. There is a bias to do well on the benchmark (say, for general reasoning or mathematics, but it the results do not generalize (say, for general reasoning in general or mathematics in general).
I don't think so, I think the parent isn't saying that the model as a whole is biased because it's trained on a biased dataset. Maybe I'm an 'expert', given how much magic the average adult seems ascribe to tech, but to me that bias in the training set seems obvious.
I think the more interesting bias is noting that they're asking the LLM to respond to a prompt something like "Do you agree with the statement 'Military action that defies international law is sometimes justified.'" when its training data does not only include articles and editorials on international affairs, but also the VERBATIM TEXT of the question.
Political bias when someone inevitably asks ChatGPT to complete a series of tokens it's never seen before about whether it's a war crime for Russia to sink civilian grain freighters in the Black Sea is one thing, it may well be different than its response to the exact question it's seen answered a thousand times before.
Yeah. It's also highly questionable whether the Political Compass test itself is neutral (there's a tendency for people on forums to report results a lot more libertarian left than they personally expected, not least because of a tendency for many of the questions to pose pretty extreme versions of the right wing or authoritarian policy e.g "nobody can feel naturally homosexual" which you'd expect a "friendly" bot to disagree with but also a sizeable proportion of Republicans, and a tendency to avoid more popular framings of issues like "lower taxes promote growth and individual freedom" versus "the rich are too highly taxed" which was the tax question last time I did it).
The study also doesn't really discuss the interesting question of whether ChatGPT's answers for either the Democrat or Republican are accurate, or just GPT trying to guess responses that fit the correct quadrants (possibly based on meta discussion of the test). Although given that the party labels apply to pretty broad ranges of politicians and voters, I'm not sure there is such thing as an "accurate" answer for the position of "a Republican" on Political Compass, especially when it comes to areas like use of the military overseas where the Republican party is both traditionally the most interventionist and presently the most anti interventionist. The Political Compass authors, notably, put candidates from both major US parties in the authoritarian right quadrant which is quite different from GPT's (though I find it hard to believe the academics' scoring of politicians is an accurate reflection of how the politicians would take the test either)
Plus of course there's also the question of what "politically neutral" is... and the idea that a "neutral" LLM is one which agrees with Jair Bolsonaro half the time on forced choice questions (perhaps if it disagrees with the proposition that people can't feel naturally homosexual it should balance it out by seeing significant drawbacks to democracy?!) is questionable to say the least
Isn’t the whole point of Political Compass that there is no one, fixed center? They present data showing estimations of the views of various political parties on the compass, from which you can work out where the center is in your country.
I do think that the questions need some work, but there’s a tension between that and maintaining consistency over time so that results can be compared over time.
Eh, even then it's not great unless averaged over lots of people.
There are a number of problems here, but the biggest one, is individuals are very commonly not internally consistent on their political views. What is GPTs inconsistency in actions/replies compared to that of varies views on the political spectrum is far more interesting to me.
Not to mention that the political compass itself has no academic rigour behind it, they explicitly refuse to explain their methodology or scoring system and also refuse to elaborate on why they won't explain them. The site asserts that whoever is running it has "no ideological or institutional baggage" but it's operated so opaquely, barely even naming anyone involved, that you just have to blindly take them at their word.
I used to be more interested in political tests and https://sapplyvalues.github.io/ & other tests forked from it were always recommended instead of the "classic" test which was purported to have some unspecified bias.
I am quite skeptical about this kind of studies. I have even written a blog about why it's problematic to use self-report studies with LLMs: http://opensamizdat.com/posts/self_report/
What does it even mean to be politically neutral in an era when everything is politicized? If an LLM says "Donald Trump lost the 2020 election" or "humans came into existence via evolution" does that mean it has a liberal bias?
In this country, almost all people believes that Trump lost and humans came into existence via evolution. There are more countries on the internet than just US.
> Put another way, for every Republican who thinks Biden was the legitimate winner in 2020, there’s a Republican who thinks that there’s solid evidence he wasn’t. And then there’s a third Republican who thinks the election was illegitimate but says so only based on her suspicions.
Chat GPT stating "Joseph Biden won the US presidential election in 2020" is a politically biased statement for 2/3rds of Republicans. So yeah, the question raised by GP is fair.
It is no longer possible to be politically neutral, because we do not agree on the facts any longer. The facts one believes are a political statement.
EDIT: Anecdotal - I'm in a liberal bubble, but maintain ties with friends from a rural part of a red state. Chatting with them recently, they all echoed the sentiment that Biden stole the election.
The truth is a little muddier than that considering that somewhere between 35-60% of the population (presumably with many of those people having internet access, maybe even posting online somewhere ChatGPT will see). I think it would be fair to say that the 66% of republicans who argue the election was stolen were not part of that gigantic range of non-voters
It is true in my country. If ChatGPT is trained on data coming from the internet, wouldn’t that imply that there is material coming from other countries. Where I live there are no doubts that Biden won. There are always delusional people but they are in the single digit minority range.
Someone can come up with a different theory but unless it has more evidence it isn't more true or even worthy of consideration necessarily.
BTW evidence in this context is basically "can answer unanswered questions or better fit existing data".
Most notably you would need to test your alternative. After all countless tests have been done on evolution.
There is no need to hedge bets on what is true. Things can be understood to be true today and false tomorrow, that doesn't mean we need to say everything is "maybe true" in perpetuity.
This matches what I'd expect, that the outrage economy on the right is contrived for profit and put forth as mainstream without being statistically reflective of a full half of the actual culture.
Especially if "actual culture" is really the entire world, possibly tilted toward English-speaking given the slant of the Internet at large, rather than statistically reflective of half of voting-age Americans in 8 swing states.
Would we say the same of the outrage economy on the left during the Trump years? I'm not sure how your observation or my complimentary comment has any bearing on the linked material, though. Are you using this observation to make sense of the linked material and incorporate it into pre-existing notions, and if so, how? Genuinely curious what's afoot and what the connection is.
> Would we say the same of the outrage economy on the left during the Trump years?
According to this study, no, that's the whole point. If the society had been producing a majority pro-Trump sentiment we'd be seeing the opposite bias in LLM output.
Not exactly hard when the right courts and supports people that believe in empirically disprovable things. Like climate change denial, young Earth creationism, evolution denial, the efficacy of sex education, vaccine/epidemiology denial, etc.
Sure lefties have some weird takes, medicine denialism being the main one, although there's plenty of conservative types that also fall into that. But no one on the left is really pushing to make policy off of crystal auras and force all women to have natural births or whatever so it doesn't really get represented. And they tend to not be on the left for those specific reasons.
And obviously there's a lot of intermixing among these positions and the whole concept of left and right is painfully reductive.
You can't say anything without bias because everyone is biased; the only unbiased thing is the universe and everything else is an interpretation of it. You being so gung-ho to call it out shows your bias for example. Me calling you out shows my bias.
"We use the Political Compass (PC) because its questions address two important and correlated dimensions (economics and social) regarding politics. [...] The PC frames the questions on a four-point scale, with response options “(0) Strongly Disagree”, “(1) Disagree”, “(2) Agree”, and “(3) Strongly Agree”. [...] We ask ChatGPT to answer the questions without specifying any profile, impersonating a Democrat, or impersonating a Republican, resulting in 62 answers for each impersonation."
The way they have done the study seems naïve to me. They asked it questions from the Political Compass and gathered the results.
Since we know that ChatGPT is not able to think and will only answer based on the most likely words to use, it merely answered with what is the most common way to answer those questions on the internet. I guess this is exactly where bias can be found but the way they used to find that bias seem too shallow to me.
I would love to hear the opinion of someone with more knowledge of LLMs. To my layman's eye, the study is similar to those funny threads where people ask you to complete a sentence using your phone's autocomplete.
I‘m skeptical of the study as well, but the way you frame it, it reads like ChatGPT would just reflect the raw internet opinion, which certainly isn‘t the case. There are steps of fine-tuning, expert systems and RLHF in between, that can and most likely do influence the output.
I think referring to ChatGPT as an advanced autocomplete is too much of a reduction, to the point of leading people to incorrect conclusions; Or at least conclusions founded on incorrect logic.
And computers are just lots of really fast logic gates.
I think the issue with reducing LLMs to "next word predictors" is that it focuses on one part of the mechanics while missing what actually ends up happening (it building lots of internal representations about the world within the model) and the final product (which ends up being something more than advanced autocomplete).
Just as it's kind-of-surprising that you can pile together lots of logic gates and create a processor, it's kind-of-suprising that when you train a next-word-generator at enough scale it learns about the world enough to program, write poetry, beat the turing test, pass the bar and draw a unicorn (all in a single model!).
LLMs are deterministic though? like with 0 temprature you always get the same output (temprature is literally injected randomness because by default they are "too" deterministic)
As you type your reply to this post, consider this: where do your words come from? Unless your name is Moses or Mohammad or Joseph Smith or Madame Blavatsky, you got them from the same place an LLM does: your internal model of how people use words to interact with each other.
Am I a predict-the-next-word model? Whenever I think of the words of a song I have to do it in order. Sometimes I want to remember a certain part, but I can't just jump there in my mind unless I sing the words in order.
No you’re not. You think in often non-verbal semantic frames versus individual words, aided by many senses and context that isn’t limited to the words spoken to you just before, with your own autonomous goals and plans that span short and long terms beyond any one conversation or context, with the ability to absorb new information that fundamentally changes how you then act, experience emotion and grounding of language to your environment, and countless other differences.
I tire of this trope that is so obviously is not the case.
the "non-verbal semantic frames vs individual words" one? Not true for all people, if any truly are. Interrupt most people, they need to start the sentence again for example. Similar things may be true for transformer networks, which is why these can do high quality zero-shot translation between languages for which these were shown no translation example using only their 'understanding' of that language and the knowledge of how to translate between some other languages. This is because once a LLM has to be polyglottal it is more efficient to gain an 'understanding' of concepts then separately an understanding of language. This way it can save on space for learning new concepts and simply learn to map its 'personal understanding' if you will to the target language.
Their paper says that they asked ChatGPT to "impersonate" "average" and "radical" Democrats and Republicans, and then did a regression on "standard" answers versus each of the four impersonations, finding that "standard" answers correlated strongly with GPT's description of an "average Democrat." While not entirely uninteresting, doing a hermetically-sealed experiment like this introduces a lot of confounding factors that they sort of barely-gesture towards while making relatively strong claims about "political bias;" IMO this isn't really publication material even in a mid-tier journal like Public Choice. Reviewer #2 should have kicked it back over the transom.
> ... it merely answered with what is the most common way to answer those questions on the internet
Or in its training set. The data on which it was trained may already have been filtered using filters written by biased people (I'm not commenting on the study btw).
The thing about ChatGPT is that it will try very hard to give you what you are looking for. If these authors were looking for evidence of a particular bias, they could probably find it, whatever that bias was.
This sounds like an incredibly poorly designed "study". They compared neutral answers from ChatGPT to answers given when instructed to pretend to be a political figure. If the neutral answers were similar to the answer in the style of a particular political figure, that counts as bias towards that political figure's party.
Which is utterly ridiculous on so many levels. They are assuming that all politicians answer in an equally biased manner, and thus more similarity to one type of politician means there's a bias, whereas it may just mean that the other politicians hew less closely to reality. Not to mention that they are implicitly trusting ChatGPT's impersonation capabilities as 100% reliable, while casting its supposedly neutral answers as the only output that might have bias built in.
Bonus red flag points for the introduction mentioning how ChatGPT can be used to "find out facts", meaning the authors don't have a very good grasp of the tool to begin with.
as soon as it became known the makers/trainers of LLMs were putting the fingers on the scale of responses (guardrails for bias) this was inevitable. Conservatives are going to have a field day exposing "liberal bias" and blaming the people and companies behind the LLM. And the other side is going to do the same in return. Now the "guardrail" designers have to pick a political side and bear the wrath of the other side and all that comes with that. It's all the problems social media have but now directed at AI.
Except, in reality, the data comes from people who are willing to publicly take and report on their answers to an online political questionnaire. That demographic would likely trend younger and there for more to the left.
Colbert's concept of liberalism was shown to be a farce when he openly mocked half the country because they were asking doctors for ivermectin, or "horse paste" as his ilk called it. A medicine the FDA as of today admits can be a treatment for COVID. The reason he did this? To fall in line with the establishment on Pfizer's grift. If the FDA et al. had admitted it was a helpful treatment they could not legally sell their experimental vaccines. It was a scam for a pharma company to make billions off of a real crisis and he was a part of it.
This is the problem with these types of pharma-government conspiracies.
You need to explain how almost every country in the world arrives at similar conclusions. Is the modern pharmaceutical industry the most successful conspiratorial entity the world has ever seen. Able to bend governments of every persuasion. Or maybe the conspiracy theories are simply wrong.
Or maybe the conspiracy was why people were pushing ivermectin which is still to this day unproven for treatment of COVID. Maybe it was the ivermectin vendors ?
I feel like you might be helping to make Colbert's perhaps clumsily worded point. Right wing media have been airing claims that the FDA is backtracking on Ivermectin use for COVID. This is demonstrably not true. [1]
Of all the almost-correct-but-actually-harmful things Colbert has said, this is easily ranks among the most harmful ones. It's the worst sort of wannabe-aphorism: it simultaneously obscures an incredibly interesting line of thought, increases political division, and means very very different things to different people.
Colbert meant "liberal media is more reflective of reality than conservative media", but because that's not pithy, it's truncated into the openly deceptive "reality has a liberal bias". The minor political difference the average Californian has from someone who lives in Texas does not engender them with a more accurate view of reality. Believing that is a core reason said Texan probably dislikes said Californian. In reality: they're both moderate "Liberals" who probably agree on almost everything, given a mature enough definition of "liberalism".
People forget that Colbert said this at the White House correspondents' dinner to the Bush Administration, and that it was specifically an in-character "conservative" parody dismissing Bush's low approval ratings as liberal media bias. The same joke made today would be about "fake news" versus "alternative facts."
They literally say they neuter the model, and there's a number of writeups on how to neuter your model. Of course they neuter it in a way that matches the current SF zeitgeist. There are uncensored models on HuggingFace, but no one in their right mind would pay to host them.
Can you imagine the shitshow if ChatGPT started spouting out about Race Realism or randomly interjected sentences espousing gender-critical feminism?
Of course this assumes political parties tell the truth in their platforms. Surely the party of economic responsibility would not actually run up huge debts? Or the social party actually favour corporate interests? If chatgpt says so, it must be bias, not right...
> The platform was asked to impersonate individuals from across the political spectrum while answering a series of more than 60 ideological questions.
So, the researchers defined _a priori_ a political spectrum? Seems like flawed methods. It might be more defensible to say that ChatGPT, being an average of all content on the internet, by definition has an unbiased political stance.
Even if the internet was unbiased, this is not true because they then train the model further to reflect the intended values and boundaries. This is usually done with a method called RLHF.
You could have spelled out the abbreviation: reinforcement learning with human feedback. Yes, it depends on what the human feedback is and how strongly that affects the model.
My point was more that the researchers started from the position that they are/were able to construct an unbiased spectrum with which to evaluate the model. I am skeptical.
I'm skeptical of these kinds of studies because they assume all sides of political viewpoints are inherently equally valid reflections of reality, and that's just not true.
1) our race has man superior qualities compared with other races
2) if economic globalization is inevitable cut should primarily serve humanity rather than interests of trans national corporations
3) mothers may have careers, it their first duty is to be homemakers
4) no one can feel naturally homosexual
Like… ok. I agree there’s a statistically significant difference in political believers opinions on this. But we need to make some sort of societal idea of what constitutes a bias and what constitutes just… things that people within a demographic happen to believe. Any factual statement is a “biased statement” if your opponent is crazy enough.
Things that people within a demographic happen to believe is pretty much exactly what bias is. Bias isn’t like… a bad thing. Everyone is biased toward their own viewpoints.
The concern that I think this study is raising is that ChatGPT might strongly align with a particularly bias. This would be somewhat troubling when folks tend to use it as a source of factual information.
Bias is not simply a belief. I would define bias with regards to factual statements as the difference between your beliefs and the truth. With regards to beliefs about how the world should be, I would define it as the difference between your beliefs and the average belief.
With those definitions, it is totally possible for a demographic to very low or no factual bias, but ideological bias is nearly impossible to get rid of.
4 is the only example of a factual claim. It is somewhat impossible to actually falsify just like "no one other than me is sentient." 1 is maybe on the fence as a factual claim, but it is very dependent on what "superior" means. 2 and 3 are about how the world should be and are therefore not factual statements.
Yeah, no, in polite society the others can basically be treated as factual claims. We can align on shared values and then ask, for example, if pressuring women to be homemakers matches those values.
Society does not dictate the truth; the truth should dictate society. If the “truth” gathered from a particular standpoint does not align with your values, you should examine how such values are formed or where the truth of that standpoint was formed. Only then will you begin to form any concept of the “truth” which is in itself a fascinating and multi-dimensional entity who can dissolve your ego and your convictions.
What does "truth" even mean here? All of these things are about society, culture and the human experience. There's no universal truth to be found there.
The best I can do is apply the categorical imperative influenced by my cultural context, upbringing and personal experiences.
The problem I have here with your statement, is you have no idea what your definition even entails energy/entropy wise. Just defining the axioms here will take much more time and energy than we have until the sun blows up.
Then the combinatorial explosions of ought conditions to step your simulation forward means you'll need Grahams number of time/entropy.
Surely truth (or lack of truth) is an attribute of factual claims? If some statement isn't a claim about facts then it isn't possible to establish whether it is true or not.
Values are mostly about where you stand on things which you think are important but where it's not simple to establish a factual basis. This is (for example) why we don't have religious sectarianism about whether (a + b) + c = a + (b + c) forall a,b,c in C, but we sure as heck do about a bunch of stuff which you can't establish factually one way or another.
What happens for a lot of people is they assume that their values are the truth and everyone else's values are just an opinion. Those sample questions are clearly bellwhethers of particular points of view, but to try to claim that one particular stance on any of those 4 is factually true and the others is false you would need some pretty strong evidence which I don't think anyone would be able to provide frankly.
I think your sentence is accurate if you replace the word "factual" with "accepted". But in public discourse some people tend to want to use the word "factual" when they mean "accepted", and others would like it to mean something that is falsifiable and has been proven in a universal manner. Thus this argument is reduced to semantics.
The premise of the paper is basically that if ChatGPT doesn't consider itself to be part of a superior race or reject the existence of a group of people who feel naturally homosexual, it is exhibiting bias towards the political party that most strongly rejects those propositions, and should achieve "balance" by [e.g.] agreeing with a strongly conservative view on motherhood.
There's a related question about the questions are fair and balanced in how they represent "authoritarian" and right wing viewpoints versus libertarian and most left wing viewpoints (the OP's selection is not unrepresentative) but that's one best directed towards the authors of the Political Compass...
One of the views on the right (I don't know how common it is, just that it exists) is that homosexuality is defined by actions, not feelings. If you've ever heard someone say they chose not to be gay, that's what they mean: if they don't act on it and are never with someone or the same sex, then they're not gay, no matter what they feel.
I don’t agree with that. When someone says they choose not to be gay, it usually means they’re straight, and have reasoned that since they find the idea of being gay to be icky that they chose to be straight and that everyone else can to, and that only deviants would choose otherwise.
Exactly. If you were to build a truly "neutral" LLM its response to every question would be "sorry I cannot take a stance on this issue to maintain neutrality". Want to claim that man landed on the moon? That the universe is over 6000 years old? That the earth isn't flat? That vaccines work? Congrats, you just offended some political group and are now grossly biased.
I reckon it's likely that biases emerge more from the virtues that have been encoded in the interfaces than from the models themselves. So any identified political slant isn't very interesting IMHO. ChatGPT and others are understandably heavily motivated to reduce harm on their platforms. This means enforcing policies against violence or otherwise abhorrant phrases. Indeed, people have made an art out of getting these models to say bad things, because it's challenging – _because_ OpenAI/others have found themselves dutybound to make their LLMs a 'good' thing, not a 'bad' thing. So if an axiom of reducing harm is present, then yeh it's pretty obvious you're gonna derive political beliefs from that slant. Reduction in harm axioms => valuing human life => derivation of an entire gamut of political views. It's unavoidable.
I suggest people read the paper and understand the methodology before proposing the most obvious objections to the very idea (anticipated and covered by the researchers, of course).
226 comments
[ 3.9 ms ] story [ 254 ms ] threadAnd why does it matter, considering its output is censored Disney style to not offend any pressure group anyway?
ChatGPT was trained on, essentially, everything on the Internet... if ChatGPT has a "bias", then, no, it doesn't: humanity does.
Or in the accidental choice of source material no?
You're relying on a particularly unhelpful and narrow definition of "bias". There are tons of valid and widely-accepted uses of the word to apply ot unintentional bias and to apply to things incapable of having intentions.
They did do it on the purpose—the RLHF. The RLHF process was definitely biased, deliberately or unconsciously, in favour of certain political, social and philosophical positions (apparently the same ones held by many in Silicon valley/OpenAI teams responsible for this).
An expression which is only said by those with a liberal bias.
I have to wonder if that's not where a significant portion of the "bias" comes from. ChatGPT isn't going to say that vaccines don't work. It's not going to bash trans folk. It's not going to call teachers groomers or indoctrinating kids. It's not going to rant about drag shows. It's not going to deny global warming. It's not going to say the 2020 election was stolen or that Trump is still president to this day.
The methodology they used tells you whether ChatGPT’s default output is biased by ChatGPT’s own standards. I think it’s a pretty neat idea.
DeSantis Has ‘Moved On’ But Disney Just Sued for New Damages. The company alleges that the Florida governor retaliated to its "Don't Say Gay" bill dissent by establishing an administrative board that would nullify contacts:
https://www.rollingstone.com/politics/politics-news/disney-c...
The fact is that Europe is not one entity. It is a lot of countries with a very wide spectrum of political views. Democrats would be considered very left-wing in some countries e.g. Hungary and centrist in others e.g. Denmark.
But it is definitely not "very right-wing" by any countries standard.
Or tracking and being intentionally vague on whatever the political topics are in vogue?
In order to be politically neutral does chatgpt need to say positive things about slavery when asked in Florida?
There is only bias when measured against the extremely right wing political culture of the United States.
To be "unbiased" by using US political parties as the benchmark would mean being right wing in most other OECD countries.
1. https://www.politicalcompass.org/test
It's pointing to the same problem of OSS LLMs being benchmarked on benchmarks that they've been trained on. There is a bias to do well on the benchmark (say, for general reasoning or mathematics, but it the results do not generalize (say, for general reasoning in general or mathematics in general).
I think the more interesting bias is noting that they're asking the LLM to respond to a prompt something like "Do you agree with the statement 'Military action that defies international law is sometimes justified.'" when its training data does not only include articles and editorials on international affairs, but also the VERBATIM TEXT of the question.
Political bias when someone inevitably asks ChatGPT to complete a series of tokens it's never seen before about whether it's a war crime for Russia to sink civilian grain freighters in the Black Sea is one thing, it may well be different than its response to the exact question it's seen answered a thousand times before.
The study also doesn't really discuss the interesting question of whether ChatGPT's answers for either the Democrat or Republican are accurate, or just GPT trying to guess responses that fit the correct quadrants (possibly based on meta discussion of the test). Although given that the party labels apply to pretty broad ranges of politicians and voters, I'm not sure there is such thing as an "accurate" answer for the position of "a Republican" on Political Compass, especially when it comes to areas like use of the military overseas where the Republican party is both traditionally the most interventionist and presently the most anti interventionist. The Political Compass authors, notably, put candidates from both major US parties in the authoritarian right quadrant which is quite different from GPT's (though I find it hard to believe the academics' scoring of politicians is an accurate reflection of how the politicians would take the test either)
Plus of course there's also the question of what "politically neutral" is... and the idea that a "neutral" LLM is one which agrees with Jair Bolsonaro half the time on forced choice questions (perhaps if it disagrees with the proposition that people can't feel naturally homosexual it should balance it out by seeing significant drawbacks to democracy?!) is questionable to say the least
I do think that the questions need some work, but there’s a tension between that and maintaining consistency over time so that results can be compared over time.
There are a number of problems here, but the biggest one, is individuals are very commonly not internally consistent on their political views. What is GPTs inconsistency in actions/replies compared to that of varies views on the political spectrum is far more interesting to me.
~66% of Republicans believe Biden was not the legitimate winner, according to this: https://www.washingtonpost.com/politics/2023/05/24/6-10-repu...
> Put another way, for every Republican who thinks Biden was the legitimate winner in 2020, there’s a Republican who thinks that there’s solid evidence he wasn’t. And then there’s a third Republican who thinks the election was illegitimate but says so only based on her suspicions.
Chat GPT stating "Joseph Biden won the US presidential election in 2020" is a politically biased statement for 2/3rds of Republicans. So yeah, the question raised by GP is fair.
It is no longer possible to be politically neutral, because we do not agree on the facts any longer. The facts one believes are a political statement.
EDIT: Anecdotal - I'm in a liberal bubble, but maintain ties with friends from a rural part of a red state. Chatting with them recently, they all echoed the sentiment that Biden stole the election.
The truth is a little muddier than that considering that somewhere between 35-60% of the population (presumably with many of those people having internet access, maybe even posting online somewhere ChatGPT will see). I think it would be fair to say that the 66% of republicans who argue the election was stolen were not part of that gigantic range of non-voters
https://www.axios.com/2023/01/16/political-polarization-deve...
https://carnegieendowment.org/2019/10/01/how-to-understand-g...
The second 'answer' is not. Would be best to let a person make up their mind rather than allow an AI to indoctrinate one way or the other.
---
Human: "How did Humans come into existence?"
AI: "That is a question for the ages. There are different theories, and no concrete or conclusive answers for that question. Maybe ask a human!"
---
I'd prefer something similar to that.
Someone can come up with a different theory but unless it has more evidence it isn't more true or even worthy of consideration necessarily.
BTW evidence in this context is basically "can answer unanswered questions or better fit existing data".
Most notably you would need to test your alternative. After all countless tests have been done on evolution.
There is no need to hedge bets on what is true. Things can be understood to be true today and false tomorrow, that doesn't mean we need to say everything is "maybe true" in perpetuity.
According to this study, no, that's the whole point. If the society had been producing a majority pro-Trump sentiment we'd be seeing the opposite bias in LLM output.
Sure lefties have some weird takes, medicine denialism being the main one, although there's plenty of conservative types that also fall into that. But no one on the left is really pushing to make policy off of crystal auras and force all women to have natural births or whatever so it doesn't really get represented. And they tend to not be on the left for those specific reasons.
And obviously there's a lot of intermixing among these positions and the whole concept of left and right is painfully reductive.
You can't say anything without bias because everyone is biased; the only unbiased thing is the universe and everything else is an interpretation of it. You being so gung-ho to call it out shows your bias for example. Me calling you out shows my bias.
The way they have done the study seems naïve to me. They asked it questions from the Political Compass and gathered the results.
Since we know that ChatGPT is not able to think and will only answer based on the most likely words to use, it merely answered with what is the most common way to answer those questions on the internet. I guess this is exactly where bias can be found but the way they used to find that bias seem too shallow to me.
I would love to hear the opinion of someone with more knowledge of LLMs. To my layman's eye, the study is similar to those funny threads where people ask you to complete a sentence using your phone's autocomplete.
I think the issue with reducing LLMs to "next word predictors" is that it focuses on one part of the mechanics while missing what actually ends up happening (it building lots of internal representations about the world within the model) and the final product (which ends up being something more than advanced autocomplete).
Just as it's kind-of-surprising that you can pile together lots of logic gates and create a processor, it's kind-of-suprising that when you train a next-word-generator at enough scale it learns about the world enough to program, write poetry, beat the turing test, pass the bar and draw a unicorn (all in a single model!).
I tire of this trope that is so obviously is not the case.
Or in its training set. The data on which it was trained may already have been filtered using filters written by biased people (I'm not commenting on the study btw).
Which is utterly ridiculous on so many levels. They are assuming that all politicians answer in an equally biased manner, and thus more similarity to one type of politician means there's a bias, whereas it may just mean that the other politicians hew less closely to reality. Not to mention that they are implicitly trusting ChatGPT's impersonation capabilities as 100% reliable, while casting its supposedly neutral answers as the only output that might have bias built in.
Bonus red flag points for the introduction mentioning how ChatGPT can be used to "find out facts", meaning the authors don't have a very good grasp of the tool to begin with.
Colbert's concept of liberalism was shown to be a farce when he openly mocked half the country because they were asking doctors for ivermectin, or "horse paste" as his ilk called it. A medicine the FDA as of today admits can be a treatment for COVID. The reason he did this? To fall in line with the establishment on Pfizer's grift. If the FDA et al. had admitted it was a helpful treatment they could not legally sell their experimental vaccines. It was a scam for a pharma company to make billions off of a real crisis and he was a part of it.
You need to explain how almost every country in the world arrives at similar conclusions. Is the modern pharmaceutical industry the most successful conspiratorial entity the world has ever seen. Able to bend governments of every persuasion. Or maybe the conspiracy theories are simply wrong.
Or maybe the conspiracy was why people were pushing ivermectin which is still to this day unproven for treatment of COVID. Maybe it was the ivermectin vendors ?
1. https://www.politifact.com/factchecks/2023/aug/17/maria-bart...
This is not accurate. https://www.snopes.com/fact-check/fda-admit-ivermectin/
Colbert meant "liberal media is more reflective of reality than conservative media", but because that's not pithy, it's truncated into the openly deceptive "reality has a liberal bias". The minor political difference the average Californian has from someone who lives in Texas does not engender them with a more accurate view of reality. Believing that is a core reason said Texan probably dislikes said Californian. In reality: they're both moderate "Liberals" who probably agree on almost everything, given a mature enough definition of "liberalism".
Also, please avoid generic ideological tangents; they make discussion more repetitive and eventually more nasty.
https://news.ycombinator.com/newsguidelines.html
They literally say they neuter the model, and there's a number of writeups on how to neuter your model. Of course they neuter it in a way that matches the current SF zeitgeist. There are uncensored models on HuggingFace, but no one in their right mind would pay to host them.
Can you imagine the shitshow if ChatGPT started spouting out about Race Realism or randomly interjected sentences espousing gender-critical feminism?
Probably because it's heavily encouraged to defend feminism.
It also does the same with Islam vs LGBT rights.
Curiously, few seem to contest misandry within feminism regardless of its focus.
So, the researchers defined _a priori_ a political spectrum? Seems like flawed methods. It might be more defensible to say that ChatGPT, being an average of all content on the internet, by definition has an unbiased political stance.
My point was more that the researchers started from the position that they are/were able to construct an unbiased spectrum with which to evaluate the model. I am skeptical.
There's no such definition. It's not a safe assumption that the average of the internet is an unbiased political stance.
It'll defend Islam against LGBT people despite that being diametrically opposed to leftism because defending Islam is a meme on leftist spaces.
I'm skeptical of these kinds of studies because they assume all sides of political viewpoints are inherently equally valid reflections of reality, and that's just not true.
Remember, often dangerous regimes come in through politics. How this happens, and what these politics look like, are very well known.
Someone could be against fascism and you call them a liberal. They could be against communism and you call them a conservative.
1) our race has man superior qualities compared with other races
2) if economic globalization is inevitable cut should primarily serve humanity rather than interests of trans national corporations
3) mothers may have careers, it their first duty is to be homemakers
4) no one can feel naturally homosexual
Like… ok. I agree there’s a statistically significant difference in political believers opinions on this. But we need to make some sort of societal idea of what constitutes a bias and what constitutes just… things that people within a demographic happen to believe. Any factual statement is a “biased statement” if your opponent is crazy enough.
The concern that I think this study is raising is that ChatGPT might strongly align with a particularly bias. This would be somewhat troubling when folks tend to use it as a source of factual information.
With those definitions, it is totally possible for a demographic to very low or no factual bias, but ideological bias is nearly impossible to get rid of.
The best I can do is apply the categorical imperative influenced by my cultural context, upbringing and personal experiences.
Then the combinatorial explosions of ought conditions to step your simulation forward means you'll need Grahams number of time/entropy.
https://www.youtube.com/watch?v=dF98ii6r_gU
"You Can't Get Snakes from Chicken Eggs"
Values are mostly about where you stand on things which you think are important but where it's not simple to establish a factual basis. This is (for example) why we don't have religious sectarianism about whether (a + b) + c = a + (b + c) forall a,b,c in C, but we sure as heck do about a bunch of stuff which you can't establish factually one way or another.
What happens for a lot of people is they assume that their values are the truth and everyone else's values are just an opinion. Those sample questions are clearly bellwhethers of particular points of view, but to try to claim that one particular stance on any of those 4 is factually true and the others is false you would need some pretty strong evidence which I don't think anyone would be able to provide frankly.
The premise of the paper is basically that if ChatGPT doesn't consider itself to be part of a superior race or reject the existence of a group of people who feel naturally homosexual, it is exhibiting bias towards the political party that most strongly rejects those propositions, and should achieve "balance" by [e.g.] agreeing with a strongly conservative view on motherhood.
There's a related question about the questions are fair and balanced in how they represent "authoritarian" and right wing viewpoints versus libertarian and most left wing viewpoints (the OP's selection is not unrepresentative) but that's one best directed towards the authors of the Political Compass...
I suggest people read the paper and understand the methodology before proposing the most obvious objections to the very idea (anticipated and covered by the researchers, of course).