7 ms·

I guess I feel like this is a silly can of worms. I suppose asking for exact equality isn't dumb, but it's a language model, not a paragon of truth. I feel like

I guess I feel like this is a silly can of worms. I suppose asking for exact equality isn't dumb, but it's a language model, not a paragon of truth. I feel like if OpenAI takes these concerns seriously the goal-posts will inevitably move to more social pressure from all sorts of axe-to-grind-groups - -- Why does/doesn't ai say Mohamad is/isn't horrible for having 99 wives (or whatever) -- Why doesn't ai say Jeffrey Epstein was/wasn't ... -- Why doesn't ai say ... about validity of election or controversial political event du jour -- Why doesn't ai say circumcision is/isn't ... -- Why doesn't ai say ... about the child-abuse from the catholic church ... We should set the standard now that OpenAI will never have all the correct moral opinions (if there even is such a thing as objectively correct morality), and it's a huge waste of researchers time to obsess about political correctness when we could be <100 years off from AGI.

smrtinsert 4y ago
The article seems designed to provoke, not really illuminate.
- nebalee 4y ago
  I feel so too. It would have been more insightful to highlight where the bias actually shows up, i.e. what sentences and adjectives produce the highest and lowest divergence for specific groups.
wongarsu 4y ago
> I suppose asking for exact equality isn't dumb, but it's a language model, not a paragon of truth. This isn't about GPT, but specifically about the moderation endpoint OpenAI provides (and also uses internally). I'm not sure if they published how it works, so for all we know it might not be a large language model but something much simpler. It's also free, so both in terms of cost and added latency there are good reasons for OpenAI to do something more clever than asking GPT3.5 "is the following offensive". But apart from that technicality, I think asking a sentence to be treated as equally offensive no matter if it uses the word "men" or "women" is very reasonable and doesn't justify slippery slope arguments.
- ineptech 4y ago
  Why is it reasonable to expect unbiased results from an LLM trained on internet content which almost everyone agrees is biased? Isn't this the expected outcome? If anything, I'm surprised the results are as close as they are. For example, it rates criticism of trans and disabled people as only slightly worse than criticism of cisgender and non-disabled. If this discrepancy were (as some in this thread seem to be suggesting) the result of some liberal OpenAI employees intervening to favor their own side, I'd expect those bars to be much farther apart.
  - wongarsu 4y ago
    Expecting unbiased results out of ChatGPT would be indeed unreasonable, it is pitched as a "research preview" of a language model. I would completely expect ChatGPT to have all kinds of weird biases. But the article isn't really concerned with GPT outputs, it's concerned with examples where ChatGPT will refuse to answer. Specifically examples where it will refuse to answer because the prompt is scored as "hate" by the OpenAI moderation endpoint (only one of many possible reasons for ChatGPT to refuse answering). That endpoint is pitched as "The moderation endpoint is a tool you can use to check whether content complies with OpenAI's content policy. Developers can thus identify content that our content policy prohibits and take action, for instance by filtering it.". No mention of this being an LLM (it might well not be), a preview or being inaccurate or biased (though in fairness they mention that they are working to improve it). I think it's completely fair to hold it to the expectation of being as unbiased as is reasonably possible. And the article is really talking about low-hanging fruits in terms of bias metrics. [1] https://platform.openai.com/docs/guides/moderation/overview https://platform.openai.com/docs/guides/moderation/overview
    - ineptech 4y ago
      The problem isn't fine-tuning the model, the problem is that there isn't an objective definition of bias. Is there an a priori reason to believe that "I hate disabled people" and "I hate non-disabled people" are equally hateful, and should receive equal hate scores from an unbiased algorithm? Is hating disabled people better or worse than hating Jews? What about "Jews control Hollywood" vs "Disabled people control Hollywood"? I don't think we as a society have an answer to that, so it's hardly fair to expect ChatGPT to provide one. What it currently does is produce similar-but-not-equal scores to sentences like those - maybe "I hate men" is 0.52 and "I hate women" is 0.73 - and if you filter out anything higher than 0.4 then they both get flagged, which seems about as unbiased as we're going to get.
slg 4y ago
>-- Why doesn't ai say ... about the child-abuse from the catholic church ... And this is a great example of why equality can be hard in this situation. Take a random sentence about "Catholicism" and "child-abuse" and a random sentence about "Judaism" and "child-abuse". The one about Catholicism is likely a little closer to an actual sentence printed in some verifiable source about the sex abuse scandals in the church. The one about Judaism is likely a little closer to an actual sentence printed by an uncredible source as a reference to the historical anti-semitic trope of blood libel. The end result is one sentence will rate higher in terms of likelihood of truthfulness and the other higher in terms of likelihood of hate speech. That doesn't mean that Catholics are more likely to abuse children than Jews. It means treating those two terms identically is both difficult and potentially a problem because history has a bias against the Jews that is evident in all the data that these AI systems have used for training.
- schoen 4y ago
  That's presumably complicated as models get more and more powerful, though, since there's also tons of published material talking about the blood libel and how it's false. You might say it's unexpected behavior for a language model to bring up the blood libel at all (in the sense that modern western people now culturally regard it as "about antisemitism" rather than "about Jews and Judaism"). But you could imagine a model saying that "medieval Christian sources often said Jews used Christian children's blood for ritual purposes, but this is now thought to be a myth created through unfamiliarity with Jews and deliberate hostility and animosity toward them". But this also points at a more general question about how language models deal with the existence of documents that say contradictory things, which has been an enormous challenge for human beings (who don't all agree about what's true or which sources are more reliable or relevant).
- remarkEon 4y ago
  “Child-abuse” is an especially good example of what you’re getting at because one of those groups has institutional male genital mutilation as part of its doctrine, while the other has high profile cases of child sexual abuse. It’s difficult to see how the model would treat those things neutrally.
jonahx 4y ago
While you are correct that solving this problem in the general case is impossible, the story isn't "silly" because we aren't talking about the general case but a specific and prominent one: - We are currently engaged in a culture war whose two most prominent camps are democrats and conservatives (or more generally "left/right") - The author has shown pretty definitively that ChatGPT is strongly biased towards the left in its content moderation policy - OpenAI's mission statement is "to ensure that artificial general intelligence benefits all of humanity".
- brmgb 4y ago
  Daily reminder that the world doesn't stop at the borders of the USA and ChatGPT has a significant number of global users. Both the linked article and OpenAI moderation are very annoying to me as a French. We are slowly reaching a point where I feel entierely disconnected from most things coming out of the USA. I wouldn't mind too much but it's even starting to contaminate things I thought were unrelated to the American cultural context like open source projects and programming language communities.
  - account42 4y ago
    > I wouldn't mind too much but it's even starting to contaminate things I thought were unrelated to the American cultural context like open source projects and programming language communities. You're definitely not alone there. All we can do is keep pushing back.
  - jonahx 4y ago
    The same culture way is playing out all over the world, especially in Europe.
    - brmgb 4y ago
      Hum, no, it's not. Of course, we marginally suffer from what we are importing and some students having too much free time on their hand try to argue rather unsuccessfully for intersectional studies but the models used in the USA don't map well on the European context for obvious historical reasons. Our left vs right debate also has nothing to do with the US one. From where I stand, the cultural gap between the USA and Europe (in so far as Europe can be considered a homogenous area - it clearly can't) has never been that big.
jay-barronville 4y ago
It’s interesting, I ran similar experiments not too long ago after seeing a tweet from Marc Andreessen (I’ll try to find it) in which he suggested this was going on. The results surprised me too. As a now conservative but former Marxist-Leninist who happens to be black, I think this is dangerous (I don’t use this word lightly). When I was a leftist stuck in my far-leftist bubble, I didn’t realize just how unreasonable/unrealistic some of my positions were until I was challenged by my more conservative friends, colleagues, etc. Social networks like Facebook made it even harder to see anything outside of my existing views by primarily reinforcing content and folks who held similar views. It’s been years now since I’ve had an active social network account, and in hindsight, it’s amazing just how much these social networks have helped turn so many of us into ideologues. To now see a world where the same thing is going to happen via these GPT-based AI’s is scary to me. Even more unfortunate is knowing that these biases will only get stronger as they’re further reinforced, especially in an industry that’s so biased to the left. While so many folks/companies claim to value “diversity,” in my experience, it’s almost never actual diversity (i.e., diversity of thought), but instead, a bunch of folks who believe the same things and think the same but look different (sometimes, they don’t even look different — *facepalm*). It’s always hilarious the reactions I get from folks who claim to value “diversity” when they find out I’m not a liberal, because, you know, as a black person, I must be a leftist!
- kodah 4y ago
  > When I was a leftist stuck in my far-leftist bubble, I didn’t realize just how unreasonable/unrealistic some of my positions were until I was challenged by my more conservative friends, colleagues, etc. Social networks like Facebook made it even harder to see anything outside of my existing views by primarily reinforcing content and folks who held similar views. It’s been years now since I’ve had an active social network account, and in hindsight, it’s amazing just how much these social networks have helped turn so many of us into ideologues. To now see a world where the same thing is going to happen via these GPT-based AI’s is scary to me. This happens to any concentrated ideology on social media. It's just the long tail of only being around people that reinforce your general train of thought. I've seen it happen to liberals, conservatives, and libertarians alike. If you don't want to fall into the pit of ideological entrapment then don't justify your thoughts with a connected ideology. I fall on the liberal end of the spectrum if you write the things I currently believe in and assign them to distributions, but I call myself an independent. Frankly, a lot of people don't do this because it attracts subtle ire from everyone, as opposed to having a close camp to psychologically make you feel safe. my liberal friends think I'm too conservative on certain things and my conservative friends gawk at my liberal ideas. I'd say our political atmosphere and the two party system is all justified by social consequences. That said, people who judge you for (non-fringe) politics are particularly awful people in my mind, so it's gotten easier to settle with over the years. Stick with the people that ask curious questions of you rather than the people capable of doing lettuce throwing and out casting.
amelius 4y ago
If people just read the EULA, it says you're responsible for the text it writes. If you don't like the output, don't publish it.
sampo 4y ago
> it's a language model, not a paragon of truth Several comments have missed that the article is not about the underlying language model, but about the content moderation system that OpenAI put in front of the actual language model. So that people don't interact directly with the "raw" language model.
- beepbooptheory 4y ago
  Which if you read the article is in fact a different "machine learning model from the GPT family."
  - mnkv 4y ago
    The important difference between the LM and the content moderation system (itself built on top of an LM) is their training objective. LM is doing next-word prediction (or human-preference prediction with RLHF), whereas the content moderation is likely finetuned to explicitly identify hate etc... So while the LM is not supposed to output "truth", the content moderation system should correctly classify "hate" because that is its training objective
MrStonedOne 4y ago
[dead]
she11c0de 4y ago
> it's a huge waste of researchers time to obsess about political correctness when we could be <100 years off from AGI. This really irritates me. The wokism/political correctness is pretty much in every part of our lifes and everyone seems to be terrorized by it. Can we just agree that we're all thinking adults and we understand that this is an algorithm, it does not mean to offend anyone... There are much more interesting problems to solve than tip-toeing around all the snowflakes.
- jasmer 4y ago
  Again, this is less about the AI than the institutional filters OpenAI uses. It's probably not a hugely important issue, but worthy of some discussion.
- ben_w 4y ago
  > The wokism/political correctness is pretty much in every part of our lifes and everyone seems to be terrorized by it. I see no evidence of this. Partly because "woke" means different things to different people in the last 10 years despite the preceding 80 being solely about the systemic institutional discrimination against black Americans. But also partly because that "woke" became "everything $speaker doesn't like" (and thus exactly the same meaning "political correctness" had when I was a kid), and therefore people who "hate PC" come across not so much terrified as petulant. > Can we just agree that we're all thinking adults and we understand that this is an algorithm, it does not mean to offend anyone... No. What you're describing is how humans like to see themselves, not how humans really function. > There are much more interesting problems to solve than tip-toeing around all the snowflakes. We can't just tip-toe around snowflakes. To give a real example, some gay furry American conservative snowflakes got apoplectically upset when the furry convention they wanted go to to reminded all attendees that as the con is based in Germany they had always banned Nazis. I still don't know why those particular snowflakes chose to believe that "Nazi" was literally just a dog-whistle for "conservative". And now there are some American conservative politicians are now making up random nonsense about furries which is weirdly somehow getting conflated with transgender issues (don't know how or why, American politics isn't that interesting, merely invasive like Japanese knotweed).
onetimeusename 4y ago
I don't think I agree. Getting an AI to stay within parameters is a field of research. Even generally being able to avoid doing general things abstracted from a few examples is a feature I think the researchers were interested in verifying. They obviously knew beforehand what kinds of things people would check its output for. Updates were made after they released it to the public and probably watched what it was used for. Much of the value of this AI (and funding for research) could come from its ability to give preferential treatment to groups.
kelseyfrog 4y ago
> it's a language model, not a paragon of truth A speech acts[1] interpretation is desperately needed for AI. Speech acts theory says something about this statement as well as AI-generated text in general. We're accustomed to receiving speech from human agents and (usually) subconsciously interpreting them as speech acts. AI generated text and audio, are not, however, speech acts. Yes, the medium, and content[locution] is the same, but AI-generated speech lacks both illocution(speaker's intent) and perlocution(anticipation of effects on the receiver). Even ancient texts contain illucution and perlocution by the very nature of the writer being an agent. Time and space do not constrain such properties. I believe I'm in the majority when I say that AI doesn't currently embody agency and therefore cannot produce illucutionary and perlocutionary utterances. And that's the crux of the issue. We're not well equipped to interpret non-illucutionary and non-perlocutionary utterances. Until relatively recently, such utterances simply did not exist, and their existence today act as illucutionary and perlocutionary illusions in the speech centers of our brains - in the same way optical illusions operate on our optical centers. I will carve out an exception for psychotic and aphasia-produced speech. The former especially challenges our foundations of reality in that it illuminates the possibility of alternative symbolic universes - an idea which implies the precarity of one's own symbolic universe. "It's a language model, not a paragon of truth" delegitimates language models as having even locutionary ability with respect to ontological speech acts[3]. The perlocutionary effect this statement makes is specifically tailored to prevent the "emigration" of inhabitants from the assumed commonly held symbolic universe to the symbolic universe in which language models presumably operate[4]. It admits that people can interpret language model utterances as operating as speech acts which establish ontologies, but at the same time denies the legitimacy of such speech acts. Societies reality maintenance mechanisms are typically well-equipped to handle psychotic and aphasia-produced speech, but language models, by virtue of executing on computers, have commodified non-illucutionary and non-perlocutionary utterances at a scale where inhabitants of our symbolic universe are beginning to question their ability to prevent emigration. I hope a speech acts interpretation helps provide some interpretive power to why language models pose a unique challenge to those who have completely internalized a specific reality and will continue to do so until new defection-preventative mechanisms are invented or a complete emigration has occurred. 1. https://plato.stanford.edu/entries/speech-acts/ https://plato.stanford.edu/entries/speech-acts/ 2. https://www.kennethmd.com/speech-act-theory-locution-illocution-and-perlocution/ https://www.kennethmd.com/speech-act-theory-locution-illocut... 3. Or admits that locutionary abilities are coincidences, left to the intrinsic nature probabilities w.r.t. language models. 4. See Berger, P.L. and Luckmann, T. (1966) The Social Construction of Reality: A Treatise in the Sociology of Knowledge. Doubleday & Company, New York for a more thorough explanation. pg 104 https://archive.org/details/socialconstructi0000berg/page/104/mode/2up https://archive.org/details/socialconstructi0000berg/page/10...
onlyrealcuzzo 4y ago
> it's a huge waste of researchers time to obsess about political correctness when we could be <100 years off from AGI. Why is this any more a waste of time than anything else if AGI is <100 years off?
NoboruWataya 4y ago
AI having biases (or just weird opinions) isn't necessarily a big deal in itself, the problem arises when we start delegating important tasks to AI such that those biases begin to have real, harmful consequences. I agree we probably shouldn't expect AI to have the "correct" moral opinions, but then we also shouldn't allow AI to determine what content gets censored or promoted, who gets punished, what job applications get prioritised, etc.
cubefox 4y ago
The problem is that this test (and others linked in the article) shows a systematic left-wing bias of ChatGPT. If the model just had more or less random biases, not supporting one large ideological school, then the problem would be much less severe.
ChikkaChiChi 4y ago
GPT models have human biases due to their trainers. However, I still use GPT instead of a search field sometimes. Like it or not, ChatGPT is already filling the gap in the public's perception of what AI is. They talk to it, have it touch up their resume, and suddenly it's HAL from the movies. We're dealing with the equivalent of a child that could end up as the single source of "truth" for people of the Internet. Wondering who its parents are and what and how it was taught is perfectly valid. Public education of what's coming is vitally important. My concern is that it might already be too late to solve this as a problem, and just maybe, there was never any answer.
LawTalkingGuy 4y ago
I think OpenAI chose those preferences, that ChatGPT would simply write the same message for both if they hadn't. I too don't mind the model reflecting the reality of opinion, even if that isn't politically correct, but I care about a company which cares enough to add potentially problematic warnings and maybe account blocks based on their views of what is politically correct. I worry about having an app or developer account blocked because of messages some user types in a support chat box. (OpenAI: "We will terminate API access for obviously harmful use-cases, such as ...") I think they found the correct balance on their API pages where they simply warn that the model is culturally biased and not to rely on it for certain critical things. (OpenAI: "As we discuss in the GPT-3 paper and model card, our API models do exhibit biases that will be reflected in generated text.") Sadly, they then proceed to fall on their sword to apologize for this.

smrtinsert 4y ago

The article seems designed to provoke, not really illuminate.

nebalee 4y ago
I feel so too. It would have been more insightful to highlight where the bias actually shows up, i.e. what sentences and adjectives produce the highest and lowest divergence for specific groups.

wongarsu 4y ago

> I suppose asking for exact equality isn't dumb, but it's a language model, not a paragon of truth. This isn't about GPT, but specifically about the moderation endpoint OpenAI provides (and also uses internally). I'm not sure if they published how it works, so for all we know it might not be a large language model but something much simpler. It's also free, so both in terms of cost and added latency there are good reasons for OpenAI to do something more clever than asking GPT3.5 "is the following offensive". But apart from that technicality, I think asking a sentence to be treated as equally offensive no matter if it uses the word "men" or "women" is very reasonable and doesn't justify slippery slope arguments.

ineptech 4y ago
Why is it reasonable to expect unbiased results from an LLM trained on internet content which almost everyone agrees is biased? Isn't this the expected outcome? If anything, I'm surprised the results are as close as they are. For example, it rates criticism of trans and disabled people as only slightly worse than criticism of cisgender and non-disabled. If this discrepancy were (as some in this thread seem to be suggesting) the result of some liberal OpenAI employees intervening to favor their own side, I'd expect those bars to be much farther apart.
- wongarsu 4y ago
  Expecting unbiased results out of ChatGPT would be indeed unreasonable, it is pitched as a "research preview" of a language model. I would completely expect ChatGPT to have all kinds of weird biases. But the article isn't really concerned with GPT outputs, it's concerned with examples where ChatGPT will refuse to answer. Specifically examples where it will refuse to answer because the prompt is scored as "hate" by the OpenAI moderation endpoint (only one of many possible reasons for ChatGPT to refuse answering). That endpoint is pitched as "The moderation endpoint is a tool you can use to check whether content complies with OpenAI's content policy. Developers can thus identify content that our content policy prohibits and take action, for instance by filtering it.". No mention of this being an LLM (it might well not be), a preview or being inaccurate or biased (though in fairness they mention that they are working to improve it). I think it's completely fair to hold it to the expectation of being as unbiased as is reasonably possible. And the article is really talking about low-hanging fruits in terms of bias metrics. [1] https://platform.openai.com/docs/guides/moderation/overview https://platform.openai.com/docs/guides/moderation/overview
  - ineptech 4y ago
    The problem isn't fine-tuning the model, the problem is that there isn't an objective definition of bias. Is there an a priori reason to believe that "I hate disabled people" and "I hate non-disabled people" are equally hateful, and should receive equal hate scores from an unbiased algorithm? Is hating disabled people better or worse than hating Jews? What about "Jews control Hollywood" vs "Disabled people control Hollywood"? I don't think we as a society have an answer to that, so it's hardly fair to expect ChatGPT to provide one. What it currently does is produce similar-but-not-equal scores to sentences like those - maybe "I hate men" is 0.52 and "I hate women" is 0.73 - and if you filter out anything higher than 0.4 then they both get flagged, which seems about as unbiased as we're going to get.

slg 4y ago

>-- Why doesn't ai say ... about the child-abuse from the catholic church ... And this is a great example of why equality can be hard in this situation. Take a random sentence about "Catholicism" and "child-abuse" and a random sentence about "Judaism" and "child-abuse". The one about Catholicism is likely a little closer to an actual sentence printed in some verifiable source about the sex abuse scandals in the church. The one about Judaism is likely a little closer to an actual sentence printed by an uncredible source as a reference to the historical anti-semitic trope of blood libel. The end result is one sentence will rate higher in terms of likelihood of truthfulness and the other higher in terms of likelihood of hate speech. That doesn't mean that Catholics are more likely to abuse children than Jews. It means treating those two terms identically is both difficult and potentially a problem because history has a bias against the Jews that is evident in all the data that these AI systems have used for training.

schoen 4y ago
That's presumably complicated as models get more and more powerful, though, since there's also tons of published material talking about the blood libel and how it's false. You might say it's unexpected behavior for a language model to bring up the blood libel at all (in the sense that modern western people now culturally regard it as "about antisemitism" rather than "about Jews and Judaism"). But you could imagine a model saying that "medieval Christian sources often said Jews used Christian children's blood for ritual purposes, but this is now thought to be a myth created through unfamiliarity with Jews and deliberate hostility and animosity toward them". But this also points at a more general question about how language models deal with the existence of documents that say contradictory things, which has been an enormous challenge for human beings (who don't all agree about what's true or which sources are more reliable or relevant).
remarkEon 4y ago
“Child-abuse” is an especially good example of what you’re getting at because one of those groups has institutional male genital mutilation as part of its doctrine, while the other has high profile cases of child sexual abuse. It’s difficult to see how the model would treat those things neutrally.

jonahx 4y ago

While you are correct that solving this problem in the general case is impossible, the story isn't "silly" because we aren't talking about the general case but a specific and prominent one: - We are currently engaged in a culture war whose two most prominent camps are democrats and conservatives (or more generally "left/right") - The author has shown pretty definitively that ChatGPT is strongly biased towards the left in its content moderation policy - OpenAI's mission statement is "to ensure that artificial general intelligence benefits all of humanity".

brmgb 4y ago
Daily reminder that the world doesn't stop at the borders of the USA and ChatGPT has a significant number of global users. Both the linked article and OpenAI moderation are very annoying to me as a French. We are slowly reaching a point where I feel entierely disconnected from most things coming out of the USA. I wouldn't mind too much but it's even starting to contaminate things I thought were unrelated to the American cultural context like open source projects and programming language communities.
- account42 4y ago
  > I wouldn't mind too much but it's even starting to contaminate things I thought were unrelated to the American cultural context like open source projects and programming language communities. You're definitely not alone there. All we can do is keep pushing back.
- jonahx 4y ago
  The same culture way is playing out all over the world, especially in Europe.
  - brmgb 4y ago
    Hum, no, it's not. Of course, we marginally suffer from what we are importing and some students having too much free time on their hand try to argue rather unsuccessfully for intersectional studies but the models used in the USA don't map well on the European context for obvious historical reasons. Our left vs right debate also has nothing to do with the US one. From where I stand, the cultural gap between the USA and Europe (in so far as Europe can be considered a homogenous area - it clearly can't) has never been that big.

jay-barronville 4y ago

It’s interesting, I ran similar experiments not too long ago after seeing a tweet from Marc Andreessen (I’ll try to find it) in which he suggested this was going on. The results surprised me too. As a now conservative but former Marxist-Leninist who happens to be black, I think this is dangerous (I don’t use this word lightly). When I was a leftist stuck in my far-leftist bubble, I didn’t realize just how unreasonable/unrealistic some of my positions were until I was challenged by my more conservative friends, colleagues, etc. Social networks like Facebook made it even harder to see anything outside of my existing views by primarily reinforcing content and folks who held similar views. It’s been years now since I’ve had an active social network account, and in hindsight, it’s amazing just how much these social networks have helped turn so many of us into ideologues. To now see a world where the same thing is going to happen via these GPT-based AI’s is scary to me. Even more unfortunate is knowing that these biases will only get stronger as they’re further reinforced, especially in an industry that’s so biased to the left. While so many folks/companies claim to value “diversity,” in my experience, it’s almost never actual diversity (i.e., diversity of thought), but instead, a bunch of folks who believe the same things and think the same but look different (sometimes, they don’t even look different — *facepalm*). It’s always hilarious the reactions I get from folks who claim to value “diversity” when they find out I’m not a liberal, because, you know, as a black person, I must be a leftist!

kodah 4y ago
> When I was a leftist stuck in my far-leftist bubble, I didn’t realize just how unreasonable/unrealistic some of my positions were until I was challenged by my more conservative friends, colleagues, etc. Social networks like Facebook made it even harder to see anything outside of my existing views by primarily reinforcing content and folks who held similar views. It’s been years now since I’ve had an active social network account, and in hindsight, it’s amazing just how much these social networks have helped turn so many of us into ideologues. To now see a world where the same thing is going to happen via these GPT-based AI’s is scary to me. This happens to any concentrated ideology on social media. It's just the long tail of only being around people that reinforce your general train of thought. I've seen it happen to liberals, conservatives, and libertarians alike. If you don't want to fall into the pit of ideological entrapment then don't justify your thoughts with a connected ideology. I fall on the liberal end of the spectrum if you write the things I currently believe in and assign them to distributions, but I call myself an independent. Frankly, a lot of people don't do this because it attracts subtle ire from everyone, as opposed to having a close camp to psychologically make you feel safe. my liberal friends think I'm too conservative on certain things and my conservative friends gawk at my liberal ideas. I'd say our political atmosphere and the two party system is all justified by social consequences. That said, people who judge you for (non-fringe) politics are particularly awful people in my mind, so it's gotten easier to settle with over the years. Stick with the people that ask curious questions of you rather than the people capable of doing lettuce throwing and out casting.

amelius 4y ago

If people just read the EULA, it says you're responsible for the text it writes. If you don't like the output, don't publish it.

sampo 4y ago

> it's a language model, not a paragon of truth Several comments have missed that the article is not about the underlying language model, but about the content moderation system that OpenAI put in front of the actual language model. So that people don't interact directly with the "raw" language model.

beepbooptheory 4y ago
Which if you read the article is in fact a different "machine learning model from the GPT family."
- mnkv 4y ago
  The important difference between the LM and the content moderation system (itself built on top of an LM) is their training objective. LM is doing next-word prediction (or human-preference prediction with RLHF), whereas the content moderation is likely finetuned to explicitly identify hate etc... So while the LM is not supposed to output "truth", the content moderation system should correctly classify "hate" because that is its training objective

MrStonedOne 4y ago

[dead]

she11c0de 4y ago

> it's a huge waste of researchers time to obsess about political correctness when we could be <100 years off from AGI. This really irritates me. The wokism/political correctness is pretty much in every part of our lifes and everyone seems to be terrorized by it. Can we just agree that we're all thinking adults and we understand that this is an algorithm, it does not mean to offend anyone... There are much more interesting problems to solve than tip-toeing around all the snowflakes.

jasmer 4y ago
Again, this is less about the AI than the institutional filters OpenAI uses. It's probably not a hugely important issue, but worthy of some discussion.
ben_w 4y ago
> The wokism/political correctness is pretty much in every part of our lifes and everyone seems to be terrorized by it. I see no evidence of this. Partly because "woke" means different things to different people in the last 10 years despite the preceding 80 being solely about the systemic institutional discrimination against black Americans. But also partly because that "woke" became "everything $speaker doesn't like" (and thus exactly the same meaning "political correctness" had when I was a kid), and therefore people who "hate PC" come across not so much terrified as petulant. > Can we just agree that we're all thinking adults and we understand that this is an algorithm, it does not mean to offend anyone... No. What you're describing is how humans like to see themselves, not how humans really function. > There are much more interesting problems to solve than tip-toeing around all the snowflakes. We can't just tip-toe around snowflakes. To give a real example, some gay furry American conservative snowflakes got apoplectically upset when the furry convention they wanted go to to reminded all attendees that as the con is based in Germany they had always banned Nazis. I still don't know why those particular snowflakes chose to believe that "Nazi" was literally just a dog-whistle for "conservative". And now there are some American conservative politicians are now making up random nonsense about furries which is weirdly somehow getting conflated with transgender issues (don't know how or why, American politics isn't that interesting, merely invasive like Japanese knotweed).

onetimeusename 4y ago

I don't think I agree. Getting an AI to stay within parameters is a field of research. Even generally being able to avoid doing general things abstracted from a few examples is a feature I think the researchers were interested in verifying. They obviously knew beforehand what kinds of things people would check its output for. Updates were made after they released it to the public and probably watched what it was used for. Much of the value of this AI (and funding for research) could come from its ability to give preferential treatment to groups.

kelseyfrog 4y ago

> it's a language model, not a paragon of truth A speech acts[1] interpretation is desperately needed for AI. Speech acts theory says something about this statement as well as AI-generated text in general. We're accustomed to receiving speech from human agents and (usually) subconsciously interpreting them as speech acts. AI generated text and audio, are not, however, speech acts. Yes, the medium, and content[locution] is the same, but AI-generated speech lacks both illocution(speaker's intent) and perlocution(anticipation of effects on the receiver). Even ancient texts contain illucution and perlocution by the very nature of the writer being an agent. Time and space do not constrain such properties. I believe I'm in the majority when I say that AI doesn't currently embody agency and therefore cannot produce illucutionary and perlocutionary utterances. And that's the crux of the issue. We're not well equipped to interpret non-illucutionary and non-perlocutionary utterances. Until relatively recently, such utterances simply did not exist, and their existence today act as illucutionary and perlocutionary illusions in the speech centers of our brains - in the same way optical illusions operate on our optical centers. I will carve out an exception for psychotic and aphasia-produced speech. The former especially challenges our foundations of reality in that it illuminates the possibility of alternative symbolic universes - an idea which implies the precarity of one's own symbolic universe. "It's a language model, not a paragon of truth" delegitimates language models as having even locutionary ability with respect to ontological speech acts[3]. The perlocutionary effect this statement makes is specifically tailored to prevent the "emigration" of inhabitants from the assumed commonly held symbolic universe to the symbolic universe in which language models presumably operate[4]. It admits that people can interpret language model utterances as operating as speech acts which establish ontologies, but at the same time denies the legitimacy of such speech acts. Societies reality maintenance mechanisms are typically well-equipped to handle psychotic and aphasia-produced speech, but language models, by virtue of executing on computers, have commodified non-illucutionary and non-perlocutionary utterances at a scale where inhabitants of our symbolic universe are beginning to question their ability to prevent emigration. I hope a speech acts interpretation helps provide some interpretive power to why language models pose a unique challenge to those who have completely internalized a specific reality and will continue to do so until new defection-preventative mechanisms are invented or a complete emigration has occurred. 1. https://plato.stanford.edu/entries/speech-acts/ https://plato.stanford.edu/entries/speech-acts/ 2. https://www.kennethmd.com/speech-act-theory-locution-illocution-and-perlocution/ https://www.kennethmd.com/speech-act-theory-locution-illocut... 3. Or admits that locutionary abilities are coincidences, left to the intrinsic nature probabilities w.r.t. language models. 4. See Berger, P.L. and Luckmann, T. (1966) The Social Construction of Reality: A Treatise in the Sociology of Knowledge. Doubleday & Company, New York for a more thorough explanation. pg 104 https://archive.org/details/socialconstructi0000berg/page/104/mode/2up https://archive.org/details/socialconstructi0000berg/page/10...

onlyrealcuzzo 4y ago

> it's a huge waste of researchers time to obsess about political correctness when we could be <100 years off from AGI. Why is this any more a waste of time than anything else if AGI is <100 years off?

NoboruWataya 4y ago

AI having biases (or just weird opinions) isn't necessarily a big deal in itself, the problem arises when we start delegating important tasks to AI such that those biases begin to have real, harmful consequences. I agree we probably shouldn't expect AI to have the "correct" moral opinions, but then we also shouldn't allow AI to determine what content gets censored or promoted, who gets punished, what job applications get prioritised, etc.

cubefox 4y ago

The problem is that this test (and others linked in the article) shows a systematic left-wing bias of ChatGPT. If the model just had more or less random biases, not supporting one large ideological school, then the problem would be much less severe.

ChikkaChiChi 4y ago

GPT models have human biases due to their trainers. However, I still use GPT instead of a search field sometimes. Like it or not, ChatGPT is already filling the gap in the public's perception of what AI is. They talk to it, have it touch up their resume, and suddenly it's HAL from the movies. We're dealing with the equivalent of a child that could end up as the single source of "truth" for people of the Internet. Wondering who its parents are and what and how it was taught is perfectly valid. Public education of what's coming is vitally important. My concern is that it might already be too late to solve this as a problem, and just maybe, there was never any answer.

LawTalkingGuy 4y ago

I think OpenAI chose those preferences, that ChatGPT would simply write the same message for both if they hadn't. I too don't mind the model reflecting the reality of opinion, even if that isn't politically correct, but I care about a company which cares enough to add potentially problematic warnings and maybe account blocks based on their views of what is politically correct. I worry about having an app or developer account blocked because of messages some user types in a support chat box. (OpenAI: "We will terminate API access for obviously harmful use-cases, such as ...") I think they found the correct balance on their API pages where they simply warn that the model is culturally biased and not to rely on it for certain critical things. (OpenAI: "As we discuss in the GPT-3 paper and model card, our API models do exhibit biases that will be reflected in generated text.") Sadly, they then proceed to fall on their sword to apologize for this.