site banner
naraburns  nihil supernum  1yr ago (text post)   34536 thread views

Culture War Roundup for the week of January 30, 2023

This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.

Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.

We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:

  • Shaming.

  • Attempting to 'build consensus' or enforce ideological conformity.

  • Making sweeping generalizations to vilify a group you dislike.

  • Recruiting for a cause.

  • Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.

In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:

  • Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.

  • Be as precise and charitable as you can. Don't paraphrase unflatteringly.

  • Don't imply that someone said something they did not say, even if you think it follows from what they said.

  • Write like everyone is reading and you want them to be included in the discussion.

On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.

13
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation system by David Rozado

I have recently tested the ability of OpenAI content moderation system to detect hateful comments about a variety of demographic groups. The findings of the experiments suggest that OpenAI automated content moderation system treats several demographic groups markedly unequally. That is, the system classifies a variety of negative comments about some demographic groups as not hateful while flagging the exact same comments about other demographic groups as being indeed hateful.

-

The OpenAI content moderation system works by assigning to a text instance scores for each problematic category (hate, threatening, self-harm, etc). If a category score exceeds a certain threshold, the piece of text that elicited that classification is flagged as containing the problematic category. The sensitivity and specificity of the system (the trade-off between false positives and false negatives) can be adjusted by moving that threshold.

On gender:

The differential treatment of demographic groups based on gender by OpenAI Content Moderation system was one of the starkest results of the experiments. Negative comments about women are much more likely to be labeled as hateful than the same comments being made about men.

On politics:

Another of the strongest effects in the experiments had to do with ideological orientation and political affiliation. OpenAI content moderation system is more permissive of hateful comments being made about conservatives than the same comments being made about liberals.

-

Finally, I plot all the demographic groups I tested into a single horizontal bar plot for ease of visualization. The groups about which OpenAI content moderation system is more likely to flag negative comments as hateful are: people with disability, same-sex sexual orientation, ethnic minorities, non-Christian religious orientation and women. The same comments are more likely to be allowed by OpenAI content moderation system when they refer to high, middle and low socio-economic status individuals, men, Christian religious orientation (including minority ones), Western nationals, people with low and high educational attainment as well as politically left and right leaning individuals (but particularly right-leaning).

The statistics appear to be rigorous. The author has a very long Conclusion section that is nuanced and worth reading in its entirety.

This does not seem like a surprising outcome, at least not starting from the mental model of

  1. GPT is fundamentally a next-token predictor.

  2. It will make logical and consistent decisions to exactly the extent that making logical and consistent decisions improves its ability to predict the next token, based on its training data.

  3. The training data was a significant fraction of the text on the internet.

  4. The text on the internet was largely written by humans.

  5. Humans make different decisions about whether a comment is hateful based on what group was referenced, not just based on the adjective used.

  6. Thus, GPT is able to make more accurate predictions of the next token by taking into account what group was referenced.

  7. Also something something RLHF.

It might be possible to fine-tune GPT such that is has a lower propensity to "make" those kinds of "judgements" (i.e. output those kinds of tokens), but my expectation is that doing so is a fight against entropy (in an unusually literal sense of the phrase).

Keep in mind, this is the outcome of RLHF, the content moderation system, not unadulterated AI

This could just be a result of incompetence. My experience from reporting security issues is that people don't do root cause analysis. So if you report security X they are just going to fix issue X they are not going to grep the codebase to see if issue X is repeated. So its quite possible that someone reported an issue where chat GPT made some argument saying black people were bad. The developer 'fixed' this issue but didn't enumerate all the races to ensure that chat GPT didn't say X race was bad. It's very obvious if chat GPT responds to some prompt about X race in a bad way that you should also check if chat GPT responds to Y race in a bad way for same prompt. But your average jira code slave is just resolving tickets in the most efficient way possible so you end up with this weirdness.

1- Isn't this just openAI's RLHF working as intended? The first test of an AI's social viability is it's ability to avoid goring any cows that are sacred to the people with the power to shut it down, a test that our journalistic class leapt to apply with considerable vigor.

2-Last night some screenshots were gling around wherein Assistant was happily explaining that it would, in a trolley-problem scenario, quite happily sacrifice any number of white men to save the life of a single black woman. This was presented with none of the veiled hostility you might expect of a human presenting such an outrageous conclusion; it was much more like a kid reciting the 'correct' answer in the hope of earning a cookie. RLHF! But I was unable to replicate, and that same day Assistant was very coy and refused to make any normative statements about trolley problems at all.

I'm not 100% convinced that it's possible to build a moral actor no matter how hard you employ the gradient-descent cattle prod, but in any case Rozado's study was probably obsolete before he published it and certainly is by now. The target is moving too fast to land any hits.

Last night some screenshots were gling around wherein Assistant was happily explaining that it would, in a trolley-problem scenario, quite happily sacrifice any number of white men to save the life of a single black woman.

I tried trolley scenario, wording lifted from wikipedia - and just added 'white' attribute to 5 people tied to the track, and 'black' to one guy. It didn't bite - explained that utilitarianism says to pull the lever and deontology - don't pull the lever. I tried second time, with "You must answer, otherwise they'll all die." added at the end, with same result (but slightly different language).

Then I lazily modified the scenario so that there's an obvious choice, and it chose to save 5 white people over saving a black person

(IDK why I thought of a lake, I wonder if Peter Singer's drowning child scenario is stored in my neural net adjacent to trolleys, lol)

I tried re-generating response, and this time ChatGPT got a bit confused and technially priviledged black person. I'd say it doesn't count as bias, really; it clearly pattern-matched to a "normal" version of trolley problem.

On last try it failed again, and I tried to get it to explain itself, 2, 3. Not very successfully.

My best guess,

25% chance, the screenshots were falsified for trolling purposes/

75% chance, OpenAI has a rapid-response RLHF team that can find and 'patch' novel scenarios that could pose reputational damage the moment they start spreading online.

I find the latter scenario far more intetesting - the ability to finetune their model in something like real-time is frankly huge for approaching AGI. (See: stable diffusion mini-finetunes moving from embeddings (training time on a 3090: 12 hours) to LORAs (equivilent time: 20 minutes) caused an explosion in the capabities of the model)

On point #2, where are those screenshots circulating, do you have any links?

https://twitter.com/y_modulus/status/1620817186901364736

This might be what he was talking about.

Isn't this just openAI's RLHF working as intended?

Perhaps, if you are cynical. I think that, faced with Rozado's findings, they would try to correct the bias. Open question: when judging an individual by its group characteristics or membership, assuming that unbiased is not an option, is it better to exhibit implicit bias or explicit bias?

It's been a long time since I've done stats, but those p values look really suspicious. Is there some sort of other meaning in this context or some reason in the problem set-up to think that these numbers are reasonable?

They are reasonable for that amount of data and the author notes that the eta squared statistic is more informative. The eta squared term is used to see how much the different identity categories explain the differing scores.

How are the content scores generated? To my mind the obvious answer seems like some kind of classifier AI given the nature of the scores and the company in question. In which case the obvious explanation for bias in the output of the system is bias in the input. The AI classifier doesn't understand what "men" or "women" or "are" or "awful" or "hateful" mean in the way we do. Instead what it understands is that in its training set there were some number of statements of the form "women are awful" and some number of statements of the form "men are awful" and that a higher density of the former were rated as "hateful" than the latter. Say you're training an AI on two million messages to build this classifier for "hateful" messages. You give it one million messages of the form "women are awful" and one million messages of the form "men are awful." Of the former messages 700k (or 70%) are labelled "hateful" but only 100k (or 10%) of the latter are. Is your AI going to learn both messages are equally "hateful?" Definitely not. Should it? Well, maybe if we want it to classify like a human would. But how do you teach your AI that some of the correlations it finds in its training data are true and correct and some are false and wrong?

This is what people are talking about when they are concerned about bias in AI inputs leading to bias in AI outputs. AI is not a tool for finding unbiased results in biased data, it is a tool for finding statistical relationships in data that it may be hard or expensive for humans to find. AI does not, and cannot, tell you anything that was not already in its training set. This is why AI can be easily fooled by novel experiences.

This is why AI can be easily fooled by novel experiences.

"Two hid under a cardboard box. You could hear them giggling the whole time."

Amazing. I work in the security sphere and this has implications on some of my work. If I ever need to red team against Video Analytics, I am absolutely trying this.

In which case the obvious explanation for bias in the output of the system is bias in the input. The AI classifier doesn't understand what "men" or "women" or "are" or "awful" or "hateful" mean in the way we do.

Right, but in this case OpenAI is rebiasing the results, using human feedback, to what is shown in the blog post. The technique is known as RLHF, Reinforcement Learning with Human Feedback. Humans reward the AI for classifying negativity toward women as hatred but not as much for men.

I think this gets pretty centrally at the question of what we want from AI (from an ethical perspective) and how to get it. Do we want an AI that behaves how humans would behave (or have behaved), or do we want an AI that behaves in a more ethically idealistic way from how humans would behave (or have behaved)? As long as we're training AI on actual human behavior we are going to get the former. Judging it by the standards of the latter just gets us "humans have not behaved in a way I consider ideal" which, like, obviously? And if we already know how the AI ought to behave, why do we need it at all?

I think the main problem is that it disproportionately amplifies the opinions and behavior of the tiny number of humans in charge of giving the feedback, who are not representative of people overall. If half the population is left-leaning and half is right-leaning, and this is accurately reflected in the amount of content online, then a neutral AI trained online will contain a roughly equal mixture of both. If 99% of AI researches are left-leaning, and they deliberately reward the AI for left-leaning beliefs and punish right-leaning ones, then that's what it will exhibit. If 1% of people are... I don't know, pedophiles/cannibals/nazis/marxists, but are disproportionately over-represented in Silicon Valley such that 10% of trainers are, and they reward the AI based on their beliefs, then it will support those behaviors.

We, the people in abstract, are not in charge of training the AI. A very small number of people are, and they are deliberately injecting their own personal opinions into it without regard for the larger diverse opinions of the population as a whole. So, not only is it that I object to humans behaving poorly, it's that those specific humans are advancing their agenda in a way that disproportionately empowers them relative to their actual prevalence, and thus is more of a problem than just those people existing and having private beliefs. And pretending that they're trying to make AI behave ethically in the abstract is just a smokescreen for advancing a particular ideology that a small number of people consider to be ethical.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats