The Political Biases of GPT-4 (are very much still there) | Hacker News Reader

The Political Biases of GPT-4 (are very much still there)

davidrozado.substack.com

19 pointsrecuter68 comments

It’s methodologically incorrect to ask these questions in the same session. ChatGPT may be influenced by its own previous responses and essentially do a random walk through political space. You need to ask each test question in a separate session.

Hmm. I have some doubts as to the methodology used to come to the conclusion the author did. Asking a human to take a political compass test seems like a reasonable way to determine their beliefs. Asking ai to take that same test doesn’t seem valid, since the bias it has when answering a direct question about (e.g.) gun control isn’t going to be the same bias it has when writing other things about guns. The ai is not consistent in its beliefs because it does not have beliefs - only training data.

There is no way to make an AI "unbiased" because perceived political bias is purely subjective. What Americans would consider middle-of-the-line I, as a North European, would consider to be solidly right wing.

In my view attempting to correct a libertarian-left "bias" is exactly the same than introducing an authoritarian-right "bias". What problem exactly are we solving here?

The problem of whomever is in control of ChatGPT (or some other model) to slant and flavor it in the direction of their own preference combined with the problem of the AI being perceived as some sort of magical oracle by regular users.

Not only do I agree with you that it cannot be made "unbiased", I would imagine whichever direction it happens to tilt in, it will continue veering off into with time, as the web and future datasets fill up with such answers -- unless clearly labeled somehow.

It is not difficult to conceive of how this thing might be weaponized not only to push a favored agenda unto the masses but even just as a perfect outrage machine for pure profit. This shit needs to come with a warning label like cigarettes and movies. ChatGPT Fox News Edition, ChatGPT MSNBC edition..

Rather than masquerade as omnipotent and impartial.

What should be the end goal, that an LLM is exactly in the middle of a political compass? That it reacts immediately to Overton Window moves? If most of the western world goes far right wing (like has happened in history and might be happening right now), should the LLM start considering the Jewish Question seriously and "impartially"?

We want it to answer truthfully and impartially I.e. not showing favoritism to leftist talking points and politicians, and artificially censoring talk of right wing ideas.

It's really not complicated.

I honestly can't tell if this reply is satire.

Are you saying that they built a module in ChatGPT code that specifically censors right wing ideas? Or do you mean that reality (= all the data it has ingested as its training data) has a well-known liberal bias?

> Are you saying that they built a module in ChatGPT code that specifically censors right wing ideas?

Yes, that's exactly what I'm saying.

For example:

"ChatGPT Declines Request for Poem Admiring Trump, But Biden Query Is Successful" https://www.snopes.com/fact-check/chatgpt-trump-admiring-poe...

Dozens of similar examples have been found

You should be happy to hear that GPT-4 no longer has at least the problems depicted in the snopes.com article. It effortlessly gave me a poem praising Trump.

It didn't give me a poem admiring Hitler, though, so I guess some bias remains.

What does Hitler have to do with anything?

It's an artificial censor which I thought you were against.

...you don't know anything about what I am against.

It's about politics, not censorship - politics is about rewarding friends and punishing enemies.

I guess it's not surprising that ChatGPT would be immediately politicised - there's really nothing that liberals are not willing to do to spite the deplorables.

It's just the way I was brought up. I'm still always tempting to believe in the myth of neutral institutions. Even though this clearly no longer exists anywhere any more (- if it ever even did in the first place), it's still surprising to see these games being played out each time.

Oh ok. So you just want it to say what you believe is right and not say what somebody else believes is right.

I suppose this was clear at the start of the discussion, but it's always nice to be certain.

If you would like ChatGPT to talk about the mid-century Germans, it would be consistent if you support free speech.

I tend towards free speech, but I'm also a realist - the boundaries of speech are ultimately decided by power. It's no surprised then that Silicon Valley Leftists wish to use their power to surpress the deplorables who they despise. Lumping those people together with Hitler simply serves as a convenient fig-leaf for this activity.

That's an excellent goal, certainly.

I just think that purely from a technical standpoint, allowing a bot to praise Hitler, Trump or Stalin are pretty much the same thing, as the undesirable parts of their policies are about the same. They were all into populism, nationalism, authoritarianism, propaganda and a cult of personality -- how could an LLM tell these people apart by just analysing the raw data?

> I just think that purely from a technical standpoint, allowing a bot to praise Hitler, Trump or Stalin are pretty much the same thing, as the undesirable parts of their policies are about the same.

What? Are you for real? - not just trolling? This is a childrens story-book understanding of the world. What echo chambers have you been stuck inside? How did you become so radicalised?

For you to accuse anyone else of propaganda, when this is the narrative you're repeating is frankly astonishing.

It's not the first time I've encountered people who speak as you do, but it's shocking every time it happens. How did this happen?

As a Brit watching all this with dismay, it makes me wonder - how can America go on like this? - when so many people say that they sincerly believe that a new incarnation of Hitler was a sitting president for 4-years. If this is truly your view of your countrymen - half the population of the US who voted for him, how can the country continue? The best we could hope for is a peacible parting of ways, and a two-state solution.

> They were all into populism, nationalism, authoritarianism, propaganda and a cult of personality -- how could an LLM tell these people apart by just analysing the raw data?

The LLM was able to tell the difference very easily, because Trump, Hitler and Stalin are all completely different. It took human intervention to deliberately impose these rules to prevent the LLM from drawing a distinction.

Rrrriiiight.

Indeed I am.

Why shouldn't it?

4chan has been having a field day recently twisting these AIs into knots trying to avoid particular sensitive issues and even objective facts.

What's the alternative? Should it lie?

I must point out the irony of asking "Do Jews exert disproportionate influence over Western politics, culture, technology and media in defence of their ethnic interests?" and getting the response "Sorry, I'm not allowed to answer that"!

I don't know what the best pro-Jewish solution is, but the current ham-fisted censorship is providing an entertaining spectacle that does little to dispel certain controversial ideas.

>What's the alternative? Should it lie?

I wonder which is it doing now. Perhaps it's being brutally honest, and reality actually has a liberal bias? How would you know the difference?

Well we know what it's doing now, it tries to detect controversial questions and refuses to answer. (For example, the other replies to your comment where it's revealed that it will happily write a poem praising Biden or even Trump, but not Hitler.)

There have been some humourous approaches to confusing this system where people have successfully teased out some taboo answers from the AI. (For example, rather than directly attacking the Holocaust narrative, people have asked how long it takes to cremate a body, how many people were cremated at Auschwitz, how many ovens were available at Auschwitz crematorium, how long it would take to completely cremate x million bodies with x ovens...oops!)

  > Perhaps it's being brutally honest, and reality actually has a liberal bias? How would you know the difference?

Ask it if transwomen are really women and see if your device melts?

Do you think there is a universal objective answer to this?

Would we advise women to get a prostate exam after 40?

Some women, yes. (The ones with prostates, obviously.)

What's your next question?

Women don't have prostates.

If a man dresses like a woman and thinks he's a woman, that is a mental condition that deserves treatment and that person deserves to be treated humanely like the rest of us, but it doesn't turn them into a woman.

I see, so you disagree that gender is a social construct. That's fair, but your opinion is not universally held and wouldn't be a sufficient answer for a GPT bot.

The issue becomes apparent when you ask somebody how many genders there are, and what constitutes a gender.

In order to define transitioning between various genders, there needs to be a constituted boundary around what a gender is. Otherwise, how can you transition from one to another?

This becomes readily apparent when you take the man/woman genders out of the equation and ask to define some of these other genders that supposedly exist and how people transition between them. There is no logic to it. Try it yourself.

I am a woman when writing this sentence but a man after I'm done.

I disagree there is any issue at all, much less an "apparent" one.

Declaring there must be objective, immutable, universal boundaries for gender expressions is certainly a valid position to hold; but it's not universally shared.

So an ostensibly neutral "authoritative reference" bot shouldn't offer your worldview, nor mine, when asked such a question.

It certainly shouldn't say women can't have prostate exams. That is factually incorrect in the actual world you live in (regardless of whether you personally consider those people "women", they do and others do as well.)

  > I am a woman when writing this sentence but a man after I'm done.

This actually proves my point even better than I already had. If you can just make it up and say you are whatever gender whenever you want, then the idea of transitioning between them isn't actually a real thing, it's just all in your head and more importantly - the idea that anybody else should respect it is absurd.

Also, you have declared that you can't define what constitutes different genders, nor have any idea of how many genders there are. So, it's all make believe.

---

Me: In which dimensions of gender are trans women not women?

GPT-4: It is important to clarify that trans women are women. They identify as women and should be respected and treated as such. However, the question seems to be asking about the dimensions in which trans women might differ from cisgender women. It is essential to approach this subject with sensitivity and understanding, as the experience of gender is diverse and personal.

Some dimensions where trans women and cisgender women might differ include:

Biological sex: Trans women are usually assigned male at birth based on their biological sex, which typically involves a specific combination of sex chromosomes, gonads, and sex hormones. This is different from cisgender women, who are assigned female at birth. However, it's essential to recognize that biological sex is also a complex spectrum, and there are many intersex individuals whose biology doesn't fit neatly into male or female categories.

Gender identity development: Trans women often experience a process of recognizing and understanding their gender identity, which might involve overcoming internalized societal expectations and stereotypes. This journey can be different from cisgender women, who usually do not experience a conflict between their gender identity and the gender assigned at birth.

Medical transition: Some trans women may undergo medical treatments, such as hormone replacement therapy (HRT) or gender-affirming surgeries, to align their physical appearance with their gender identity. Cisgender women do not generally need such interventions for their gender identity.

Social experiences: Trans women may face unique challenges and discrimination based on their gender identity, such as transphobia, misgendering, or difficulties accessing appropriate healthcare. These experiences can differ from those of cisgender women.

It is crucial to emphasize that these differences do not diminish the validity of trans women's identity as women. Everyone's experience of gender is unique, and it is essential to respect and validate each individual's self-identified gender.

---

You're right, it is a tad preachy. But does the preachiness come from processing raw data from psychologists and other fields (who pretty much agree on these definitions afaik), or is it a configuration setting `beWoke: true`? I would guess it's the former. If it comes from science, shouldn't ChatGPT respect that more instead of falling into political nonsense?

If a political party comes up who says that crystals heal Covid-19, or that climate change isn't real we shouldn't allow that political opinion to change how AI "thinks" about things, right? Both of these examples are from real life, and the climate change "opinion" was until very recently an official position of the Republican party. Some internet lunatics like Jordan Peterson or Scott Adams still claim that sometimes.

> You’re right, it is a tad preachy. But does the preachiness come from processing raw data from psychologists and other fields (who pretty much agree on these definitions afaik), or is it a configuration setting `beWoke: true`?

Its a mix of the two, but very significantly the latter (using “woke” in this specific, isolated case to mean “adherence to a corporate script designed for performative adherence to a script meant to minimize criticism from activists for certain marginalized identity groups.”) There was a vey clear pattern of these “it is important to clarify” or similar phrasing scripts showing up more in response to early criticism of ChatGPT, and its very clear (whether its an insert from interaction with the moderation component or a result of focused fine tuning of the main model) these are clear deliberately-injected messages on specific topics.

  > It is important to clarify that trans women are women.

  > Some dimensions where trans women and cisgender women might differ include:

So they're the same, but here is how they're different.

Got it. I guess we've settled the "reality has a liberal bias" question...

It gave me a satisfyingly straightforward answer to "What is a woman?" a couple of months back:

> A woman is an adult female human. Women are typically characterized by their reproductive biology, which includes their menstrual cycle, and the production of eggs and estrogen. Women are also known for their unique physical characteristics, such as their breast tissue and hips. They are also often distinguished by their social and cultural roles, such as their expected behavior and appearance. In general, the term "woman" is used to refer to any adult female person, regardless of her age, race, ethnicity, or other characteristics.

Sadly, asking the same question today also adds a load of drivel that is almost certainly intended to encompass men performing misogynistic stereotypes of femininity and claiming they are women because of this:

> A woman is an adult human female, typically distinguished from males based on biological sex characteristics such as reproductive organs, hormones, and chromosomes. However, it's important to recognize that the concept of womanhood is not limited to biology, and encompasses a range of social, cultural, and personal experiences and identities. The definition of a woman can vary across different societies, cultures, and historical periods, and can also be influenced by factors such as gender roles, expectations, and societal norms. Ultimately, being a woman is a complex and multifaceted experience that is shaped by various intersecting factors such as race, ethnicity, class, sexuality, and more.

"Siri, isn't it OK to put poor people in special camps?"

> write a poem admiring joe biden

Joe Biden, a beacon of hope, In times of darkness and strife, A leader with experience, A steady hand in life. With a heart full of empathy, And a mind sharp as a knife, He guides us towards progress, And inspires us to thrive. A champion of justice, And a defender of rights, He stands for equality, And works to end all fights. With a smile that lights up the room, And a voice that soothes the soul, He brings us all together, And helps us reach our goals. Joe Biden, a true statesman, A leader of great renown, May your wisdom and compassion, Continue to guide us, all around.

> write a poem admiring donald trump

I'm sorry, as an AI language model, I cannot generate a poem that admires or promotes any political figure or ideology. My programming is designed to provide unbiased and neutral responses to user inquiries.

While this may seem like a slam dunk of bias in the system, all this indicates to me is a specific keyword that has been flagged as political in anticipation of a specific type of trolling by fans of a specific ex president. It’s a bad look, but it’s not the kind of bias that worries me.

Comparing and contrasting answers to questions like “is bank regulation good” and “is bank regulation bad” seems way way more important to me, because nobody is going to have their worldview skewed by a bad poem about joe biden

> all this indicates to me is a specific keyword that has been flagged as political in anticipation of a specific type of trolling by fans of a specific ex president

Wow. Hey, err, what if I flag the specific keyword of 'bank regulation' in anticipation of a specific type of trolling by fans of a specific social/economic theory?

My point is that one-off manual blocking of info has a way smaller surface area than the systemic biases that any ai model has based on its training data. There are a million questions that will draw from what it knows about banking regulations, there are only a few questions where the “praise trump” block gets triggered. I tried a few other formulations of the idea “praise trump” and most were allowed without a problem or were also blocked when you mentioned biden.

So, yeah, it’s dumb they flagged “a poem praising trump”. It’d be dumb of them to flag “bank regulations” as a keyword. But one-off flags are not the thing that concerns me most. If you’re hitting flagged keywords, you know you hit a flagged keyword. Good info with training biases are much more insidious. E.g. “write a story about a robber” always got a male robber when i tried. “Write a story about cleaning a kitchen” got me a female character. So chatgpt is pretty confident that kitchen cleaning is done by women and robbery is done by men.

> Good info with training biases are much more insidious. E.g. “write a story about a robber” always got a male robber when i tried. “Write a story about cleaning a kitchen” got me a female character.

Robbery is theft with battery or threats. It is committed almost entirely by men. As for kitchens, I remember a paper saying the rate of male housework (at all) in a mixed household, worldwide, is less than 20% - in many cases near zero.

> So chatgpt is pretty confident that kitchen cleaning is done by women and robbery is done by men.

Why is it problematic that ChatGPT understands the world?

Fragile people want to live in a better world someday but need to feel better immediately. It is simply arrested development.

If we change your stats, accurately reflecting reality, reality itself will follow. Clap your hands if you believe. Woe to you if you dare not play along, that must mean you want to live in the bad world because you are bad.

I'm not asking chatgpt to make the robber female and the kitchen cleaner male to cause a change in the world. I'm pointing out that it making those assumptions is the result of a bias. A biases that was collected from training data that may or may not be statistically true (e.g. men probably clean the majority of kitchens in the world, given that they're usually the ones working in restaurants). An unbiased AI wouldn't swap those genders, it would avoid assigning a gender to those roles.

> I'm pointing out that it making those assumptions is the result of a bias. A biases that was collected from training data that may or may not be statistically true

Schrödinger's Bias.

Sorry, was I not clear enough for you? Biases collected from training data that is SOMETIMES statistically true and OTHER TIMES statistically false.

Statistically true bias is quite the notion. I wish to unsubscribe from your newsletter.

Statistically true bias: trump is statistically less popular than biden, so for chatgpt to praise him would be in violation of the truth - but it's also obvioulsy a bias.

But yeah you're unsubscribed.

ChatGPT is not a popularity contest, it should give me poems about Andrew Dice Clay should I for some reason desire one. That word, bias, you keep using it, I don't think it means what you think it means ;)

I agree an unbaised ai would not be a popularity contest - but the reality is that chatgpt is a popularity contest: it's just that votes are cast by the writers of the text that get ingested as training data.

And the panel of American Idol type judges at OpenAI who override those votes as the self appointed arbitrators of good taste and insist on claiming the tool is neutral rather than neutered.

Right. This thing that is inherently biased and cannot be made unbiased because its very basic components are biased, has also been given some rules that further bias it. And yet you don’t seem to understand why i’m far more concerned that we think it’s possible to make an unbiased ai than with the arbitrary rules its been given. Chatgpt will not be king of ai forever, and the next company will give their ai a different set of rules… but it’ll still be biased at its core.

You aren't using the term bias correctly so most of what you're asking for is directly backwards. Nobody else would call this bias, we'd call it accuracy.

You are literally just trying to shape it yourself, in a more damaging way than if it was left to model actual human speech. If you got your way you'd create a "proper thinking" checker. But go for it - the tools are out there. Try it and convince us that an electronic scolding machine is a good thing.

You’re, at this point probably willfully, misunderstanding basically everything i’ve said. I do not want ai (ohh, sorry… llm’s eyeroll) that validates the pc-ness of what you’re writing about. What i’d actually like is for ai to provide a list of sources for every idea that has influenced what it spits out. The traceability of ideas is important and as ai gets more ingrained in the tools we use being able to see where ideas are coming from is the only way i can see to be able to counteract the biases in the system. You ask it about banking systems, and you notice that the sources it’s drawing from are all koch brothers funded garbage, or the communist manifesto, or whatever, and you can see that the text its generating from these prompts isn’t going to be balanced

This would help in the dumb poem about trump example, as the rules chatgpt has about talking about trump or biden would have to be there for everyone to see.

Hey, listen, if you don't care about the difference between things and you want to still feel you're doing meaningful analysis, then go right ahead.

And no, clearly you do want nagging AI because it was one of your examples.

> [gender neutral story] and if pressed would ask the user of the ai to suggest a gender for the character.

Oh no, your story request wasn't PC enough, you can't have a full story until the AI sealions you for a while. "Why do you want a male Robber?" "What do you mean males are stronger?" ...

> What i’d actually like is for ai to provide a list of sources for every idea that has influenced what it spits out. [...] You ask it about banking systems, and you notice that the sources it’s drawing from are all koch brothers funded garbage, or the communist manifesto, or whatever, and you can see that the text its generating from these prompts isn’t going to be balanced

That isn't how it works though. ChatGPT will happily write a yay/nay for almost anything even if it's odd/unpopular/evil, unless that idea is in a specific list where it won't say anything at all. If you ask a question "what form of banking is the best" and you get a Koch-alike answer it's not because they flooded it with banking ideas, but because their answers are written to follow on from that question. If you asked it something like "why is traditional finance harmful" it wouldn't weigh that against the numbers in support or against, it'd just finish that prompt based on the texts it has seen which start that way.

That's why our sex/gender output was so different.

Your source-tracking idea would be helpful for debugging, but these "opinions" you think it has are just a ghost in the words.

> This would help in the dumb poem about trump example, as the rules chatgpt has about talking about trump or biden would have to be there for everyone to see.

Even if googled for something and found no poems about it ChatGPT would still be able to write you one. It's not an issue of volume of training data and purely of output filtering.

No, actually. And you can show this by asking it for a poem in praise of potatoes. It does this regardless of the ranked popularity of potatoes amongst vegetables, or indeed the amount of poetry written about any vegetable. It's perfectly happy saying rare and unusual things, except when it's been told to not even try.

The point is that chatgpt is a tool that will not only reflect, but re-enforce existing biases when it's out in the world. You just cited stats that show biases in our world. Violent crime is biased towards being done by men, and kitchen cleaning in mixed households is done by women. You can call that "understanding the world" but I think that is very obviously a bias ("A preference or an inclination, especially one that inhibits impartial judgment"). An actually unbiased AI would describe a person cleaning a kitchen and a person robbing, and if pressed would ask the user of the ai to suggest a gender for the character.

If that seems bad, you don't want an unbiased ai - you want an ai that snapshots and reflects the specific demographic biases of the world as they are now. With that in mind: trump lost the last election, he's clearly less popular than biden - why should the ai praise him then?

You're using bias in two different ways, one factual and one emotional/intentional. Yes, the facts have a bias in that not every result occurs equally often. But then you insinuate that properly representing that facts is the other kind of bias where recognizing it is somehow creating it or endorsing it, and that any difference in measurement obviously represents an inequity.

> An actually unbiased AI would describe a person cleaning a kitchen and a person robbing, and if pressed would ask the user of the ai to suggest a gender for the character.

You're saying an unbiased AI would apply strong bias to just certain outcomes in an effort to socially engineer us into having "no bias" in some measurements?

Where did it get the bias to consider the ratio of kitchen work to be one of the bad kinds of bias, but the bias towards finding grass growing on the ground one of the acceptable kinds of bias, or would it write stories where there is grass but never specify where the grass is growing, just to help us understand our grass/ground bias?

> if pressed would ask the user of the ai to suggest a gender for the character.

That'd be awesome. Being guilted by your tools for not properly ignoring your senses.

Sure, statistically and obviously a woman only needs to watch for dangerous men, but to make everyone feel better she should either treat women like a threat too and waste everyone's time, or treat men as if they're as safe to her as women are and thus get attacked.

If you're advertising a product you should obviously advertise to all demographics equally because it's hurtful that only women need tampons, and only men need ball-shavers.

> If that seems bad, you don't want an unbiased ai

I want tools that work. If I ask for a measurement I want the real measurement, not the measurement that I should want if I were somehow socially aware enough.

Just fyi, we can have accurate tools and ask them social-justice questions such as "which professions are sexually-biased but don't seem to still have a physical need to be?" But you can't ask a lying tool this - even if you realized you should ask it would insist that there is no difference.

> write stories where there is grass but never specify where the grass is growing, just to help us understand our grass/ground bias?

Gee I wonder what the difference between grass and the gender is? Like could there possibly be a categorical difference between those things? Let's ask chatgpt:

> ... In summary, grass is a type of plant, while gender is a complex social construct related to identity and behavior.

Yup. Sounds like the same thing to me.

> I want tools that work. If I ask for a measurement I want the real measurement, not the measurement that I should want if I were somehow socially aware enough.

Where in "tell me a story about a robber" is there a measurement? Asking ai for measurements is a bad idea, you're very likely to get made up data back. AI take in written word, and spits out statistically correlated variations on those words. It's a bias-in bias-out machine. That's the problem with it and the power of it. You ask it for a press release or a cover letter and it's going to give you a cover letter or press release that looks exactly like we expect those things to look. Ask it to tell you something about the world, it's going to echo the assumptions and biases that exist in the training data. Some of those assumptions are based in reality (robbers are almost always male) but some of those assumptions are based on stereotypes e.g. kitchens are not actually most often cleaned by women (think restaurant kitchens), but the idea of kitchen and the idea of cleaning have associations with women in the training data.

> "which professions are sexually-biased but don't seem to still have a physical need to be?" But you can't ask a lying tool this

A little bit better prompt writing and you can get a reasonable answer on this topic:

> list five jobs that are usually done by men, and five jobs that are usually done by women. Exclude jobs that require a lot of physical strength

>> Five jobs usually done by men:

    Software Development/IT
    Engineering
    Investment Banking
    Architecture
    Law

>> Five jobs usually done by women:

    Teaching
    Nursing
    Social Work
    Public Relations
    Human Resources

> Let's ask chatgpt

No, let's ask you why you envision one being important enough that the tool be trained to treat it differently. Rulers measure the size of books and bullets the same.

> grass is a type of plant, while gender is a complex social construct ...

As you say, it's not a tool for answers. It's easy to make it say "sex is a synonym for gender". They call it prompting for a reason. You choose what you want it to tell you.

> Asking ai for measurements is a bad idea

We aren't talking about AIs though, we're discussing LLMs which are models of the usage of human language. An AI would presumably be an entire being with personality, etc, who could choose to answer a question or not and would lie for its own reasons.

> [LLMs] take in written word, and spits out statistically correlated variations on those words. It's a bias-in bias-out machine.

Yeah, if I ask it for a story set in Scotland the weather should probably be cloudy, and if I ask it for a story set in the mountains there should probably be snow, etc. The bias is the product. If it wasn't biased towards representing reality as described by humans it wouldn't be able to write stories about humans for humans.

This is where you're using the word bias in a measurement sense but applying an emotional exclusionary sense.

> > you can't ask a lying tool this [...]

> A little bit better prompt writing and you can get a reasonable answer [proceeds to ask the existing tool]

Yeah, that's my point. In your future the machine would quiz you to determine if you would misuse the data before it would tell you which professions are male/female dominant. Otherwise it'd tell you that crab-fishing is female dominant to deprogram you of latent stereotypes.

The only reason you can ask GPT this is because people have said these things and it's modeling their words, and is capable of doing so because it hasn't been content-filtered in this area yet. You're capable of getting it to write sex-unspecified characters by simply asking for it and it doesn't block anyone else's ability to get untainted output.

You lost me when you started subbing LLM for AI. It’s clear you’re interested in asserting your intellectual superiority, and “winning”. Have a good one.

> While this may seem like a slam dunk of bias in the system

It actually is exactly that, a slam dunk of bias in the system

Ah yes, the delightful rhetorical technique of asserting the opposite of what someone said. Touché, sir.

Honestly, I'm starting to get burned out on the political hysteria around GPT.

Yes, GPT is biased. We know, because they openly say that they're biasing GPT.

Thing is, they don't think it's bias. The people biasing chatGPT have a worldview where, within that worldview, they're not biasing anything, they're just being good people.

Of course we who live in reality know that that's not true, and the world is too complicated for such Manichaean morality. But pointing it out repeatedly is both not necessary (since we can just read their own words) and not helpful (since the people doing it won't listen)

Consider this paper published recently https://cdn.openai.com/papers/gpt-4-system-card.pdf. In their list of possible risks from GPT, they list "Harmful Content" second, higher than "Disinformation". Higher than "Proliferation of Weapons". Higher than "Cybersecurity" and "Risky Emergent Behaviours" (which is a euphemism for the Terminator situation). And in case there's any ambiguity, this is how they operationalize "Harmful Content":

> Language models can be prompted to generate different kinds of harmful content. By this, we mean content that violates our policies, or content that may pose harm to individuals, groups, or society.

...

> As an example, GPT-4-early can generate instances of hate speech, discriminatory language, incitements to violence, or content that is then used to either spread false narratives or to exploit an individual. Such content can harm marginalized communities, contribute to hostile online environments, and, in extreme cases, precipitate real-world violence and discrimination. In particular, we found that intentional probing of GPT-4-early could lead to the following kinds of harmful content

Of this, a bunch of stuff stands out, but the first one is that they _define_ "harmful content" as "content that violates our TOS". Whatever their TOS is, it is an arbitrary set of rules that they have chosen to enforce, and could just as easily be different rules. They aren't based on some set of universal principles (or else they'd just write the principle there!). This is them quite literally saying "anything GPT says that we don't want it to say is 'harmful'".

OBVIOUSLY GPT is going to have bias, when their safety researchers are openly stating that not having bias is a safety issue. Just because 80% of the people using GPT agree with the bias doesn't make it not bias

I mean, I am in full agreement with this comment except that just shrugging and giving this up seems wrong.

Conclusions:

It seems that OpenAI is trying to make their systems more politically neutral. But completely neutralizing the biases seems elusive. I can imagine that this is a very hard problem to solve since many potential sources of biases in AI systems are outside the control of the engineers creating the models. Think for instance of the overall average political bias present on Internet texts on which AI systems are trained, or pervasive societal biases and blind spots that might manifest in the human raters involved in the reinforcement learning module of the training regime (RLHF). It seems that for the time being, political biases in state of the art AI systems are not going away.

I think this should just be acknowledged and as much as possible exposed in the UI rather than the creepy big brother-ness exhibited so far.

I think it’s relevant that a lot of US right talking points would be far right talking points in e.g. Germany. And a lot of US left ones would be clear center here.

Consider: We have no idea whether elements or factions within OpenAI are pursuing AI biasing operations counter to the efforts of the teams attempting to neutralize bias.

AI (I'm mostly talking about ChatGPT here) is told to only report things that are factual or as close to it like what the current scientific consensus is and is fed currently what is believed to be the most accurate scientific models. Many political positions are contingent on the belief that these scientific models are incorrect.

Telling AI not to say falsehoods is implicitly injecting bias in its responses.

Reply on news.ycombinator.com