site banner
ZorbaTHut  2yr ago (text post)   35549 thread views

Culture War Roundup for the week of September 19, 2022

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.

33
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

I wrote a post about de-biasing efforts in machine learning, which got a bit long, so I decided to turn it into an article instead. It's about how corporate anti-bias solutions are mostly only designed to cover their asses, and does nothing to solve the larger (actually important) issue.

(As an aside: does it still count as a "bare link" if I point to my own content, just hosted elsewhere?)

Well yes corporations just want to cover their asses from criticism, but the worst is that the criticism is so powerful that the discussion ends up shallow. I'm so annoyed that we had a cancellation of a language model that produced rap lyrics. FN Meka had a record contract which was later torn up because it was a "appropriator" using words reserved for someone with "lived experiences". It was just reported on and everyone glossed it over and forgot about. I'm like what did just happen? They cancelled a pile of math... There are slightly more deeper philosophical problems with the way it was cancelled. It cuts directly into postmodernism and its issues with treating the inauthentic the same as the authentic. Because at the end of the day the language model is just a mirror of the language used by those with the "lived experiences" and should be understood as such. https://youtube.com/watch?v=vYmfnfljvnM

My instinct is that the rap world would look down on a bunch of geeky mathletes who tried to write rap verses about being hard-ass gangsters regardless of whether they use an AI to do it or just pen and paper.

I looked into the youtube comments of FN Mekas music videos. There I found complaints of "inauthenticity" as in it wasn't very good. Which is not a surprise. The complaints highlighted in media is "robot racism" not that it sucks as rap. https://nypost.com/2022/08/24/capitol-records-severs-ties-with-virtual-rapper-fn-meka-after-outcry/

They cancelled a pile of math...

Clearly they didn't. They "canceled" actions of people for (to them) socially relevant reasons.

This kind of "they canceled math" reminds me when pirates said that "a number was banned", meaning that the sharing of an encryption key was banned (my point is independent of agreeing with that ban), or "I just used words" or "I just moved a mouse and clicked on things". Or saying "it's just pixels" if you were caught with CP.

(Another somewhat similar trick is "no person is illegal", making it seem as if someone who uses the phrase "illegal immigrant" meant that the person themselves is illegal besides being an immigrant, when clearly it means a person who immigrates illegally. The person isn't illegal, but their actions are. Similarly here, it's not the math that is canceled but an action performed through math.)

It's an annoying rhetorical tool. It's not "just" that, and the fuss isn't about that "just" part but the consequences, the context, the intent, the usage, etc. Everything is, at the end of the day, "just" something. A bomb is just some chemicals, just some molecules. Everything is just a bunch of quarks and electrons and so you can make any action sound absurd.

Or saying "it's just pixels" if you were caught with CP.

This one has an obvious full argument that it's short for, basically looking like "child molestation is bad because it (often) directly hurts children; fapping to real CP only indirectly hurts children and fapping to fictional CP doesn't hurt children at all, because the representations of children on a computer screen are not actually children but just nonsentient pixels, so the last in particular shouldn't be subject to the stigma of the first and the second is somewhere in-between". There is definitely some meaning to that word "just".

The copyright one is somewhat-further removed from the full form, which basically looks like "illegalising true and consensual information exchange is an unacceptable compromise of liberty, and copyright near-uniquely does this and indeed almost entirely consists of this". Kind of weaker, though, as it requires a fairly-strong assumption and there's at least one other example of this with pretty-broad societal support (classified information).

Sure, but my point is that saying "It's just a number! That's ridiculous to ban!" gets around the meat of the argument (any digital information can be encoded in binary and therefore as a number). It's lazy. Argue the substance.

One can be annoyed at reductionism but when people decry the object for itself, it's no longer reductionism.

"Guns aren't banned, only possessing guns is banned" is about as stupid as the converse.

The people who don't like this AI don't like it for essential reasons, not because of actions taken with it.

Also, regarding the gun example:

Stupid arguments: "guns aren't banned, just possessing guns is banned". "Banning guns makes no sense, guns are just molecules and atoms arranged in a certain configuration! You can't ban physical reality!"

Sensible arguments: "Guns should be banned because they are dangerous and people may accidentally fire them or shoot someone in the heat of the moment", "Guns should be allowed because if criminals know that people have guns, they won't break into people's homes and this makes life safer and we can all sleep better."

These are just examples.

I wonder if they would be okay with it if it (the rapper AI) was operated by black people. If yes, then it's not merely about the AI as an object or the math. You have to argue (and I don't think it's hard) that blacks don't have a monopoly on rap, it's not enough to say "hurr durr it's just math".

I wonder if they would be okay with it if it (the rapper AI) was operated by black people. If yes, then it's not merely about the AI as an object or the math. You have to argue (and I don't think it's hard) that blacks don't have a monopoly on rap, it's not enough to say "hurr durr it's just math".

So I didn't pay much attention to this controversy, mostly out of lack of interest, but also partly out of unfortunate paucity of coverage, but I think the main issue the "cancelers" had with the AI was that it was generating lyrics implying some sort of personal struggle as a black man facing police brutality, and also possibly using certain slurs that are common in rap songs produced by black people. I'm not sure if the "cancelers" would be okay with the AI if it had been operated by black people, or if they would require that the AI also be written by black people. And if the latter, does it have to be written from scratch with no use of, say, open source code created by white or Asian researchers?

Or would even that not be enough, and the mere fact that it's a computer generating these lyrics instead of a human brain belonging to a black man who had actually suffered police brutality generating them is a problem? If so, it seems reasonable to characterize that as people "canceling a pile of math;" however, it's not clear to me that this is so.

I see it totally opposite. What they are doing is precisely not about canceling the math but the social context, the use. That people made a black man imitator robot, which they consider some kind of blasphemy/taboo, like depicting the prophet for Muslims or how some tribes don't want to be photographed. It's a sacred thing for them. Just like putting some fancy Native American headdress onto some random robot or scarecrow would probably be.

You can argue against that view and that it's bad to consider such things sacred. But it's strawmanning them to pretend that they are simply arbitrarily canceling a bunch of math. They are canceling the building of software tools that impersonate black rappers, as a legitimate endeavor out of quasi religious reasons.

But it's strawmanning them to pretend that they are simply arbitrarily canceling a bunch of math. They are canceling the building of software tools that impersonate black rappers, as a legitimate endeavor out of quasi religious reasons.

No, that's not a strawman, that's a description. By your reasoning, the concept of "canceling a bunch of math" simply doesn't seem to exist; if I decided that "2+2=4" is blasphemous and demanded that all books and teachers making such statements be "canceled," I wouldn't be "canceling a bunch of math," but rather, say, "canceling the expression of certain mathematical equations out of quasi religious reasons." Mathematical equations and algorithms - whether they state some arithmetic fact or they allow the creation of some string of letters that some people find meaningful - aren't just lying around for people to discover; they are expressed and created by humans. Objecting to people expressing and creating such things is exactly what is described when something is described as "canceling a bunch of math."

If I oppose you shooting a bunch of cannon balls at my house, I'm not canceling ballistics, I'm canceling your use of ballistics in this particular way.

"Cancelling" math would be something like what the Pythagoreans did (though cancel isn't the best verb) when trying to suppress the proof/discovery that root 2 isn't rational (please let's not get bogged down in the historicity of my characterization of this particular example, it's an analogy). To consider mathematical facts themselves as dangerous or harmful. These people didn't cancel machine learning or convergence proofs or gradient descent. They cancelled an application of machine learning. If I say that a Chinese style social credit system or all encompassing surveillance infrastructure are bad or should not be implemented, I'm not cancelling the math. Just the application area.

More comments

You are annoyed at the rhetoric of my attempt of being succinct. So let me expand on that. Yes they "cancelled" actions of a group of people that own and operate a language model that was trained on existing rap lyrics. My statement was reduction on that. The thing that I'm trying to convey here is that the discussion with the ethics around input and output AI models in general is very shallow. The discussions revolve around harm caused by unpredictable output which have no actual consequence other than making someone feel bad over a generated content. Image or lyric...

The ethics discussion should be deeper than social panics du jour. It shouldn't be ethical for example to use machine learning algorithms to nudge behavior of people. e.g. presenting alcohol ads to people that has a signal in their online behavior that they are prone to become alcoholics. The other example which I saw recently on a friends youtube, ads for investment scams when he searches for how to invest in stocks(since he most likely don't have requisite knowledge to sniff out scams). Those AI applications have the potential to ruin someones life as opposed to n-word by algorithm or violent nightmare fuel from an AI painter.

Sure, this is a much better argument because now you have to address the substance of who (or what, if an AI isn't a who) is allowed to utter the n-word and whether non-blacks can utter it indirectly through the operation of AI systems or whether it causes "harm" or not.

Yes and what is the difference between an AI and CD recording? Predictability? Determinism? Am I allowed as a non-black operate a CD-player with rap album in it? To sample it and create a new piece of art?

All these questions because some people decided to equate the copy with the original. AI just mimics reality but it isn’t reality.

Try blasting the song "Nigga Nigga Nigga" as a white guy next to some wokes. I guess they'd disapprove.

(also CDs are grandfathered in, their analog predecessors were made much earlier than the current progressive morals around race).

What happens if the Spotify algorithm recommends and blasts a song with "Nigga Nigga Nigga" next to some wokes? Is that racist? Do I have an argument that "the algorithm" did it?

I could guess that if a spotify recommendation played "Gary Barts Ntu Troop - Uhuru Sasa" and "Gil Scott Heron - The revolution will not be televised" it would be deemed as appropriation.

It seems like you’re asking the model to do two contradictory things, at the same time? It can’t both be “neutral” and give you a female president half the time, and be accurate about its knights or doctors or what-have-yous. There was never a female president of the USA, so an accurate model shouldn’t generate one.

Also, I’m not sure what was wrong with the redcoats in Normandy (other than the extra arms). It’s what I thought the prompt should generate, too, and I think that’s what most people will think of.

There was never a female president of the USA, so an accurate model shouldn’t generate one.

Just an aside on this: the model is trained on images, not real life. So if there were ever representations of a female president, then it'd still be accurate, even without debiasing techniques. This could include art, woman playing the first female president in movies or TV, etc. I'm sure there have been many instances of those things. Also, if the image generation models do some sort of semantic embeddings (I have no clue if they actually do), then they may know that "president" is similar to "leader" or "politician", whatever, and can then leverage images of female leaders or politicians in figuring out what a female president might look like.

Sure, but that’s exactly the sort of “bias” that OP was talking about, with reference to e.g. medieval knights not being represented accurately. If accuracy is wanted, then no female presidents.

It seems like you’re asking the model to do two contradictory things, at the same time?

I mention that there are two approaches, targeting a politically correct world or the real one, not both at the same time!

I’m not sure what was wrong with the redcoats in Normandy

This was to contrast how ahistorical generic "knight" prompts are, akin to the AI unbiddenly putting redcoats in WWII if it was similarly bad at modern history.

I mention that there are two approaches, targeting a politically correct world or the real one, not both at the same time!

Ah, I guess we agree then!

I pointed this out years ago: there are basically two natural philosophies when it comes to avoiding bias: either it's sufficient to (e.g.) mask race as an explicit input, or it's necessary to ensure the predictions treat all races identically. If you're against both forms but keep talking about how a system isn't "fair", you need to take a long look at what, exactly, you mean by "fairness".

Also this goes well beyond machine learning -- policing, insurance, etc. A lot of political issues aren't even at the point where productive discussion is possible, because nobody is willing to state unequivocally what they mean by "fair".

mask race as an explicit input

"Unfortunately" for the machine learning case, they lack the complex internal self-censorship that humans can do to be able to pull that off. Even if you mask out inconvenient inputs like race and gender the model will likely immediately notice clusters of correlated traits that stem from that, and reconstruct the race and gender from scratch.

(A fun idea for a dystopian story element: people conspicuously purchasing items and visits places associated with a "safe" demographic like elderly Asian women, to keep the eye-of-sauron AI off their backs.)

I'm not sure I really believe "self censoring" is a coherent concept. For example, if you're predicting the probability somebody will default on a loan and white people have lower credit scores than asians people, is any model that notices assigns asians a lower probability (on average) not "self censoring"? If 99% of people who eat sushi default on their loans, is it "unfair" to penalize them?

To be frank, I've never heard a definition of "ignores race" that isn't implicitly asking for just nakedly pretending all races have equal odds of defaulting on loans, regardless of whether that's accurate or not. I think people are actually asking for this "post-hoc fairness" should be explicit about that. Instead when I say "so you're saying we should just retroactively make the outputs of the result less accurate and pretend men and women are equally likely to commit murder", the response I get is "you're just straw-manning me, I just want a model that's fair".

It's frustrating that the complaints about bias seem contradictory and/or unsatisfiable, and the people making them are unwilling or unable to elaborate. If somebody is going to criticize my model, they should give me a well-defined notion of fairness that it's actually possible for my model to meet. If their definition of fairness means "deliberately cripple your model and force banks to give out loans that are unprofitable" they need to actually own that instead of hiding behind ambiguity.

Edit: The "post hoc" solution (which I think is the only solution that meaningfully satisfies progressive demands) is:

  1. train a model that uses every variable

  2. train a model that only uses variables you want to ignore (race, sex, etc.)

  3. your predictions are model1(x) - model2(x)

Apparently actual progressives disagree, but I've never heard anyone give an alternative.

Am I right in coming away with the conclusion that your post seems to be arguing about two very different and almost opposite things?

  1. An AI's general willingness to combine things in unrealistic ways. This is the ability required to produce images of things not heavily represented in the training set, such as female presidents.

  2. An AI's ability to understand and ignore "spam" in its training data (e.g. popular depictions that mislead from "reality"), such as Marvel depictions of Norse mythological figures or faux 80s illustrations.

In a sense, these are directly opposite goals because 1 requires painting something unrealistic, and 2 requires ignoring/penalizing unrealistic outputs. I suppose the common ground is that an AI should default to painting logical/coherent/realistic things unless prompted otherwise. But even this desire is loaded - first of all, the way these AIs are trained, I think, sets them up to hopelessly fail at any measure of how "realistic" their outputs are - we humans have the advantage of our perception of reality being a distinguished input, and also by access to vastly more information crucial to understanding concepts such as causality and physical intuition. It's also clear to us whether we're seeing something real or seeing something fictional, by virtue of that fictional thing always being a subset of the reality we perceive (e.g. a still image on a display, rather than something we're seeing with our naked eyes directly).

In machine learning the aim is to distinguish signal from noise, to extract things that will generalize to future data, as opposed to merely coincidental, contingent things that appear in the training data. This however leads to deeply philosophical issues that some engineer types tend to ignore out of a feeling of superiority over humanities types. It's not clear for example how to distinguish real, fundamental correlations from mere happenstance. For example one may say that a female US president is perfectly possible, it's not a logical contradiction. The fact that we haven't had one is just a contingent fact about our timeline so far. Or perhaps that the fact that black people currently commit more crimes is not an inherent property to them but a happenstance based on external conditions and so on.

By the way, Saul Kripke just died a few days ago. He had a lot to say about necessity. I'm not convinced that the necessary vs contingent stuff (which goes back all the way to theological arguments) is actually all that meaningful but it's good to recognize that people have already discussed these sorts of things.

It's not clear for example how to distinguish real, fundamental correlations from mere happenstance.

You mean, how to distinguish correlation from causation? Isn't this exactly the domain of the scientific/empirical method, and its associated toolbag of trickery?

No, not causation. I mean correlations that will hold at test time and aren't just coincidences. Causation is a different thing.

What other options are there for a nontrivial correlation? One where both are causally connected to a common third factor?

This however leads to deeply philosophical issues that some engineer types tend to ignore out of a feeling of superiority over humanities types. It's not clear for example how to distinguish real, fundamental correlations from mere happenstance

This is precisely backwards. It's not the STEMLords declaring things racist, or biased just because they happen to have disparate impact on groups. They're also not the ones hostile to the idea of setting up experiments to find out if the connection is causal, or just a coincidence.

I'm referring to dismissals along the lines that "it's increasing the accuracy therefore it can't be unfair", which is often a very narrow minded view. Note by the way that I am certainly closer to the techie nerd type than the humanities type. My point is though that it's not so simple. And also that fairness, justice, ethics etc. are important values and they should not be conceded to be claimed by the woke as their own conceptual turf.

Maybe you train a model based on inputs age, sex and race, and you can predict some important social outcome with 76% accuracy. The conclusion shouldn't be "Great, now deploy it because it clearly captured something about reality", but to think more "why did we choose the aspects that we did? Could we use inputs that are more in control of each person and has clearer causal connection to the outcome? What happens to those that are wrongly classified? Are there substantially different subgroups who we should handle separately instead of lazily lumping them together? What is the right granularity?" etc. I'm not saying that this isn't done, but many techies like to short circuit all this and almost seem to say that if the model is more often correct than it would be by chance then its good and should be used.

All I'm saying is that this stuff is not trivial and not all of this can be easily dismissed. If you already think about these things, that's great. But I see many tech people go too much to the other end (what the woke call tech solutionism - though they overuse the term to wrangle any credence or authority or stake out of the hands of the engineer types to boss them around from their DEI offices).

Can you link an example of someone going too far the other way? As far as I can tell this whole conversation predates AI by decades, and as far as philosophy goes, it's really not that deep. It's just another take on "ice cream sales predict drownings".

(As an aside: does it still count as a "bare link" if I point to my own content, just hosted elsewhere?)

It's not ideal. I'd prefer if you at least included the first few paragraphs of text from it.

Alternatively, post it both here and on your blog, include a link to the blog at the bottom. You're welcome to advertise as long as it's the same content you're posting :)

both here and on your blog,

I wouldn't say it's a blog at all, and no advertisement was intended. It's just that GitHub is a convenient host for long-ish texts with images. The content was intended entirely for themotte; there is no other link or index to the text except the post I made above.

Hah, didn't even realize it was a Github page.

Huh. I admit the formatting is pretty nice on that.

Note that it's unlikely to be read as often through a link, and we really do want people reading it. I think this may be an argument to improving our rich-text setup, although there's definite problems with that as well. But it may still be worth doing.

I'm divided on this and will have to think on it, but you're welcome to keep it up!

I just did some corporate training about data retention, and only retaining what is absolutely necessary. Nearly every time you throw out or never collect data that is irrelevant. But one point that jumped out at me is that "throw away the data about race in our AI model" is the wrong answer as far as HR policies are concerned.

Still reading the article, but in response to your aside: I’m not sure. I have seen people do so without comment, I’ve seen them include choice quotes or a discussion question/sub statement, and I’ve seen copy-paste of the whole linked article. Reception depends on overall tone, political valence, phase of the moon, and whether or not the post sounds like JB. Personally, I favor the second option of a submission statement. Unless one is burying the lede and trying to sneak in a rule-breaking article instead of a rule-breaking post, that’s usually good enough.

You are right, they are not really trying that hard. Anybody smart enough to build bleeding-edge AI systems is smart enough to understand why if you try to predict the likelihood of a criminal repeating a crime, it will always say that black people are more likely to repeat (it's because black people are more likely to repeat). The problem is fairly hopeless, because AI's accurately that black people are more likely to commit crimes, women are for the most part uninterested in studying machine learning, and other things that true but verboten.

So their manager asks them to do something about bias, and they apply the laziest possible hack. I think this disinterest is more prominent in top-tier researchers. Low-end researchers who will never accomplish anything useful are happy to feast on the de-biasing funding teat.

There are some other niche cases, like facial recognition software not recognizing blacks. But this requires no special debiasing effort, it is simply a weakness in the system that can be addressed the same as any other weakness.

Obviously, the only possible "de-biasing" technique that can work is explicitly biasing systems against white men. If two criteria are mutually conflicting such that one group or the other must be "discriminated" against according to one criteria or the other, choose the criteria that discriminates against white men (in that order: first discriminate against whites, then discriminate against men). It is very simple.

Anybody smart enough to build bleeding-edge AI systems is smart enough to understand why if you try to predict the likelihood of a criminal repeating a crime, it will always say that black people are more likely to repeat (it's because black people are more likely to repeat).

The question is are you measuring "blackness" or are you measuring "membership in the underclass" which thanks to both accidents of history and explicit policy is comprised disproportionately of black people?

You might also be measuring membership in a group whose ancestors were less shaped by executions.

https://journals.sagepub.com/doi/pdf/10.1177/147470491501300114

It can also be that membership in the underclass is the effect not the cause.

No, it’s not obviously the only workable technique.

Half the point of this parent article was about the confusion, willful or otherwise, between the objective and the normative. Is the model supposed to be drawing like a human, or like an idealized paragon of virtue? If you ask it to draw a medieval knight (per someone else downthread) do you want it to match history or popular culture?

Predictive ML is a clear example of the objective case. Given the lack of a market for a robot which lies to you, such robots aren’t popular. Generative ML is a different story.

Train on data from the world as it is, and you’ll get relative objectivity, plus journalists will call you racist. Papering over that with prompt engineering is a hack to get more normativity for an AI which doesn’t “need” to be objective. But this suggests another strategy: don’t train it on the world as it appears to be. Train it on the world that could be.

Is creating such training data trivial? No. Does it require discriminating against anyone? Also no. Seems like a decent idea to me.

But this suggests another strategy: don’t train it on the world as it appears to be. Train it on the world that could be.

Is creating such training data trivial? No. Does it require discriminating against anyone? Also no. Seems like a decent idea to me.

Presuming that this theoretical "world that could be" is at all different from the "world as it appears to be", it absolutely requires discriminating against someone. There's just no way to bridge that difference without applying some discrimination against someone at some point in the process; otherwise we'd just end up back where we started.

Not as long as we’re talking about generative instead of predictive. For example, an AI which does everything like DALL-E except it doesn’t need prompt engineering to make diverse pictures.

Imagine a specialized net which only draws armor. We can train it on the entire internet, in which case it will draw us Renaissance artisanal plate, which has been inserted into countless historically-inaccurate pieces of media. If we instead only use a historical database of curated and tagged armors, it will end up with a better understanding of actual medieval armor. It will also be much less flexible, but it will be defended against criticisms of renaissancism.

So by your lights, what would it take for a generative AI model to be discriminatory against someone? An "AI that does everything like DALL-E except it doesn't need prompt engineering to make 'diverse' pictures" would necessarily involve discriminating against someone when creating the training set; if no such discrimination were required, then the 2 models would be identical. You seem to be saying that that doesn't count. So what would count?

I think there are a couple uses of the word "discriminatory".

First, yeah, defining a training subset is a decision to favor one group over another. Making the historically-accurate-armor AI is going to disappoint the fantasy-art users. An AI that makes everyone [insert feature here] is going to disappoint the people wanting it to be realistic. And of course there's nothing stopping the dataset curator from excluding all members of a race, or all Christians, or all women with realistic proportions; doing so would be obviously discriminatory.

On the other hand, a sentencing or profiling AI is categorically different. When a predictive machine discriminates, it is causing a disparate impact on actual people rather than on the collective. I'm struggling to find the right words...it is the difference between perpetuating an injustice vs. causing it, a sort of active vs. passive harm.

(I think there's also an argument to be made that the ill-trained AI is more obvious that an AI trained well on an evil world, and thus more compatible with exit-rights liberalism, but I'm not so confident in that...)

The OP argued that predictive AI was a zero-sum game where only reverse racism could compensate for the inbuilt discrimination. I think...there's some degree to which that is correct? It's extending that conclusion to generative AI that strikes me as wrong. I believe generative discrimination is an easier problem than predictive because it's a lesser problem. Maybe you can't make a "world that could be" generative AI without excluding anyone, but I'm reasonably sure you can make one without actively harming any individuals.

Anyway, thanks for pressing me on this. It's an interesting philosophical topic.

OK, so it sounds like based on this post that when you wrote:

Does it require discriminating against anyone?

about a theoretical generative AI model based around a world that could be (instead of the world as is), you were making a general statement about how no generative AI model could possibly discriminate against someone. Which seems like a strange thing to say about 1 specific type of generative AI model (i.e. based on a world that could be) when discussing 2 different theoretical types of models that are both generative (i.e. the other one being based on the world as it is), but technically true, I suppose.

More or less, though I wouldn't frame it as an absolute, if only because that's inviting extreme counterexamples.

The original article was writing about unimpressive debiasing in corporate generative models. When sulla responded with the assertion that this is unavoidable in a bleeding-edge AI, I thought it appropriate to point out that his examples were all predictive rather than generative. I think that really undermines his equation of discrimination with the "true but verboten." A generative AI could be wildly discriminatory, in the loose sense, without ever discriminating in the narrow, personal sense, and it is the latter which is a one-way ticket to a media circus.

If my model generates pictures of people from an even ethnic spectrum, I do not believe I am discriminating against anyone. It's not the Harvard auditions, and this isn't actual people I'm failing to generate or generating excessively.

If I'm a designer using a generative AI to fill a world of my making, I do not care for "accurate" demographic representation being baked in the model. Either I know exactly what ratios I want and include them in the prompt (and if I'm the type to want "accuracy", I already know who I want to be 13% of the generated population yet 50% of the generated criminals), or otherwise it should give me a selection of 50% male, 50% female, [for race in races generate total*race/races] American presidents.

I see the division between "factual predictors" and "diverse generators" AI models to be quite acceptable and even desirable. Make it explicit and call out those who present their model as one when it in fact has characteristics of the other.

If my model generates pictures of people from an even ethnic spectrum, I do not believe I am discriminating against anyone. It's not the Harvard auditions, and this isn't actual people I'm failing to generate or generating excessively.

By this logic, then no model is discriminating against anyone. If a model only ever returns white women for "good person" and black men for "bad person," then that's not discriminating against anyone since it's not the Harvard auditions, and this isn't actual people you're failing to generate or generating excessively. Great, looks like we can associate any race with any quality we want in our model and not worry about discriminating against anyone!

If I'm a designer using a generative AI to fill a world of my making, I do not care for "accurate" demographic representation being baked in the model.

Indeed, and notably no one seems to have any problem with such models existing. There's room for multiple types of generative AI models in this world, including ones that have uniform distributions, desired distributions, best-attempt-at-accurate distributions, or really any other arbitrary distributions of demographics.

If a model only ever returns white women for "good person" and black men for "bad person," then that's not discriminating against anyone since it's not the Harvard auditions, and this isn't actual people you're failing to generate or generating excessively.

Do you really think your example is as egregious as generating perfectly uniform selections for both "good person" and "bad person" (or "criminal" and "Harvard student", for that matter)?

Do you really think your example is as egregious as generating perfectly uniform selections for both "good person" and "bad person" (or "criminal" and "Harvard student", for that matter)?

What does being egregious have to do with this in any way? As you wrote, either way, "it's not the Harvard auditions, and this isn't actual people I'm failing to generate or generating excessively." Given that the reason that perfectly uniform selections isn't discrimination has literally nothing to do with the type of distribution and everything to do with the fact that these are generated images rather than actual people, we can change the distribution to anything we want (including my example of encoding "good person" with "white woman" and "bad person" with "black man") and still land at the same result of "no discrimination is taking place."

What does being egregious have to do with this in any way?

Pretty clear to me that the analogy you deployed was deliberately absurd.

But anyway, provided that you clearly label your Pro-White-Women, Anti-Black-Men model as such, go ahead.

More comments

This is why I distinguished between predictive and generative tools.

A robot that makes inaccurate predictions is next to useless. One that makes unrealistic art is not. Engineers trying to avoid accusations about their generative model don’t need to add discrimination, because unlike in real life, changing the training data is an option.

I would expect the “dream data” version to have all the same mainstream use cases as DALL-E or similar, except also avoiding Problematic picture generation.

Which begs the question of why we should care what journalists say, but lots of people do anyways

I don't think the companies specifically care about the journalists' opinions as a terminal goal. Rather, the (realistic) fear is that average people (potential paying customers or just people who might popularize the brand) read the articles written by the journalists and adopt those opinions and stances.

Anybody smart enough to build bleeding-edge AI systems is smart enough to understand why if you try to predict the likelihood of a criminal repeating a crime, it will always say that black people are more likely to repeat (it's because black people are more likely to repeat).

An alternative explanation is that doublethink required to simultaneously believe in the party line and in the reality required to do your job doesn't actually work very well and tends to devolve into believing in the party line only. Imagine that you're a bright young guy working on a Google's image classifier. To generate the thought that the classifier might confuse black people for apes so you must specifically check that it doesn't, you must believe that black people tend to have certain ape-like facial features. That's a very dangerous thing to believe, your woke peers would be very unamused if you just blurt it out or inexpertly wink-wink nudge-nudge your way to suggesting that you need to check for that etc. If you have a lot of wrongfact beliefs you have to watch your every word to avoid committing a social suicide. Accidentally releasing a classifier that does in fact mistake black people for apes on the other hand is relatively safe: it's not your personal fault and who could have thought and it's probably bias in the training data anyway. So in a highly ideologized environment people just naturally fail at their jobs instead of trying to maintain a bag of forbidden beliefs.

This example doesn't work - black people are dark, apes have dark fur, image classifiers often pick up on easy-to-detect features like color.

More generally I'd question how important the party line/reality conflict is - many genuine smart people believe in wokeness and will probably continue to indefinitely. E.g. OpenAI is clearly woke yet manages to put out a great product.

So their manager asks them to do something about bias, and they apply the laziest possible hack.

I actually have a different impression: most of these professional ML researchers and engineers genuinely wish they could serve up a model that provides politically correct responses, because politically correct responses are also commercially correct, and everyone wants to make money. Probably the main reason a bunch of giant and amazing Google models aren't made available to the public via API is because of the risk that they might say or display something politically incorrect, and certainly some fraction of the user base (especially tech journalists lusting after those sweet engagement metrics) will try to bait it into doing so.

So there's ample incentive to solve this problem "the right way," and the fact that so far all we see are cheap hacks and opacity is because no one knows how to solve it the right way, or even if it is solvable the right way at all, even in principle, with the technology we have today.

Part of the problem is exactly what makes these models so exciting to begin with. They can notice things, they can extrapolate from training data, they can make analogies and they can roll with out-of-sample prompts, and they develop all of these amazing abilities ex nihilo, from a largely uncontrollable black box made of inscrutable matrices gently nudged in the direction of data.

The other part of the problem is that political correctness isn't a well defined or static problem. It is a messy social problem, involving subtle adversarial factional games, sort of like fashion.

And these two halves of the problem compound with one another. It isn't enough to generate a black person one time in X -- you have to define X, you have to solve this equation for all possible identities, and you have to then translate this equation into every conceivable fact pattern that the user will (adversarially) use to challenge the model. If you want to generate a picture or story involving a policeman arresting a criminal, it is fraught whether you make the policeman white or black, whether you put him in a wheelchair or not. Should the model generate trans women? If they're visibly identifiable as trans women, are you making a minstrel caricature to further the stereotype that trans women look like men in dresses? If they aren't visibly identifiable, how is one to know they are trans at all, and that you haven't committed the deadly sin of erasure? Should black women look like white people but with a darker skin tone (and draw criticism for e.g. straightening her hair, itself a political minefield), or should you make them look recognizably phenotypically black in terms of facial features and hair (and draw criticism for reinforcing a stereotype)? If both murderers and NBA stars are disproportionately likely to be black, does the model need to recognize that murderers are bad and NBA stars are good and apply its distortion of the underlying distribution only to the bad category, i.e. return mostly white guys for criminals but mostly black guys for NBA stars? How is it to know? And when ideological opponents start to stress-test these categories and ask for a thuggish NBA player or a corrupt President, should it reverse the categories? What about middle grounds, like an "aggressive" NBA player, or a "desperate, nonviolent" criminal? We even have minor culture wars about the perceived race of robots.

I'll point out that the problem might not be so unsolvable as you describe; prompt engineering being what it is, a very thinkable (but dystopian) way some more-capable future version of DALL-E might resolve this is by adding to the prompt "and also, make sure to never portray X ethnicity negatively."

Yes, I think this sort of "prosaic alignment" solution is likely to solve all of our consternations about AI capabilities at least as well as a human intelligence could... in the long term. Eventually, you wouldn't even have to talk about portraying X ethnicity negatively, you could just say "and make it politically correct" and the AI would understand those rules better than any individual. For the time being, though, Dall-E has a hard enough time drawing a complex but coherent picture, much less enforcing its conformity with protean standards of political correctness.

Worth pointing out that OpenAI tried this sort of "prosaic alignment" approach to its so-called diversity filter. It appears to append stuff like "black male" or "hispanic female" to some proportion of prompts that it believes call for the depiction of a person. It has been vigorously panned by the community, because it has unintended negative effects on many prompts. Gwern did some sleuthing on why a complicated prompt for a picture of a cowboy at a certain angle in certain lighting etc. returned a bizarre misfire, and eventually discovered that the same prompt with "cowgirl" instead of cowboy worked flawlessly -- seemingly implicating the diversity filter in the original prompt's total failure.

Hilariously, OpenAI's approach here was discovered by asking for stuff like "A person holding a sign that says" -- and then you'd often get a picture of a sign that says "FEMALE" or "BLACK". So there's a degree to which adversarial prompt construction can overcome attempts at coercive prosaic alignment, at least using current techniques.

It also doesn't know our delicate rules about when it's socially appropriate to re-gender or trans-racialize the subject of a picture. It's weird if prompts for Princess Zelda return a black or Asian Zelda, or if prompts for George Washington return a colonial-era woman in a white wig. Maybe we'll accept that sort of thing by the time Season 10 of Bridgerton comes out, but I don't think we're there now, and it would take a pretty advanced AI to figure that stuff out.

The social-rules-about-reracialization thing is definitely a reasonable one; that's a significant issue that would result in many funny PR disasters.

On reflection vulnerability to adversarial prompt injection seems almost innate to the technology, considering both the above "person holding a sign that says " attack and also the more recent one with remote.ly.

That runs into the problem of "what counts as 'negative?'" Traits don't come with in-built value judgments; it's up to us to decide which side of some dichotomy, for example, is better. Sometimes it's blindingly obvious to all sane men which is better, but often it isn't. Many things can be cast in different lights to praise or condemn somebody depending on whether one already likes them or hate them, and so if your goal is to avoid any associations that anyone could consider negative - especially if there are motivated defenders who'd love to claim you as a prize - I think there is little, if anything, that's safe.

The details of what counts as "negative" would be determined based on the language model's own ideas of what constitutes "negative" based on its time spent with the training data. This is likely, for the most part, to align with conventional understandings of what is "negative".

Which could satisfy the proverbial Reasonable Person, perhaps, but not the proverbial Cardinal Richelieu ("If you give me six lines written by the hand of the most honest of men, I will find something in them which will hang him.") It would be nice if appealing to common sense would be all it took to deflect such attacks, but we've ceded that possibility earlier in this hypothetical from the very premise of this reputation-managing AI.

How about things like VLMs inadvertently putting out black Donald Trumps? Or more broadly, if I use a model to generate “Republican Senator”, what’s the ideal number of black or other ethnic minority faces to produce? Are we going to keep up with the liberal facade that a Senator is a Senator, regardless of political alignment, and thus we should see a diverse representation of races? Or will we instead accept that “Republican Senators are privileged white guys” and turn out a distribution of faces that supports the progressive narrative? These are points of tension within the modern left, so the only winning move is not to play. And before you suggest “just show the accurate racial distribution for a given prompt”, consider that the liberals at least still have to pretend to care about consistency, so committing to “actual truth above normative truth” as a principle is an invitation to embarrassment when the same principle is applied to other domains, eg, CEOs or nurses.

These are points of tension within the modern left, so the only winning move is not to play.

Traditionally the only winning move is to side with the most radical and potentially vengeful faction based on a more realistic version of the Basilisk/Pascal's wager theory.

I think that's one of the reasons people and institutions radicalize so quickly; there are punishments for anyone who doesn't stay ahead of the curve, but none for those who get ahead of it.

In this case there are clear incentives to start "sanitizing" Problematic output even in silly and arbitrary ways, because the value is showing that you're an accomplice. The details don't matter, as long as you throw out some shibboleths like "female-presenting nipples".

I think that's one of the reasons people and institutions radicalize so quickly; there are punishments for anyone who doesn't stay ahead of the curve, but none for those who get ahead of it.

Hence accelerationism, I think: there exist memeplexes that are fortified to the point of invulnerability from one direction but have no answer to attacks from behind. So the idea, it seems to me, goes that the way to defeat them is to attack from the unguarded direction and lure them into an untenable position ( "I dare you to step over this line!") Which I think is a pretty risky strategy, really, as the phase transition may not be as near as one would hope.

Oh, shit, I didn't know about the Black Donald Trump thing. That's hilarious.

Yeah, okay, it's a fair cop; even such a policy as I describe would result in amazing PR debacles.

Clearly the AI is a fan of the Prince song "Donald Trump (Black Version)"

My article really only covers generative models, like the recent Stable Diffusion. Controversial models like classifiers that try to evaluate how likely somebody is to commit a crime has entirely different considerations. Maybe I should have made that more clear.

Also I disagree that a "de-biased" crime model would discriminate against white men! Men commit a highly disproportionate amount of crime compared to women; any sort of adjustment you make has to adjust for that, adding a whole bunch of likelihood on women especially, probably more than the racial difference even.

I dunno, the parent comment by sulla strikes me as basically calling out a similar (though more inflammatory) situation. We have two possible meanings of the term "bias" in common use, and these two meanings are:

  1. Not faithfully representing statistical realities present in the data.

  2. Not faithfully representing the statistical outcomes that we would like to see in the data-- most commonly, which reflect reality except for not showing differences based on race or gender.

These are, of course, mutually exclusive definitions; e.g. as pointed out in your article the president of the United States should always be drawn as male using definition (1) and should half the time be female using definition (2) . Likewise, classifiers determining how likely someone is to commit a crime ALSO have to make a decision between definitions (1) and (2) while facing the complicated issue of how to avoid public controversy over admitting that these are different things.

You suggest a third, equally plausible definition (3): "Not faithfully representing the statistical outcomes present IN REAL LIFE (as opposed to just the data being trained on)."

That actually strikes me as brutally difficult, running into the same basic issues as fact checkers do now-- evaluating what is true or false in real life is really hard and intersects with political agendas in such a way as to make it even harder. And how do you even evaluate if you succeeded? Don't get me wrong, i think it's reasonably likely that some generative models will get fine-tuned on specific datasets curators will have labeled as similar to "real life" along various dimensions. But I would not anticipate that this will end up becoming the norm.

As an aside, I think it makes a lot of sense that fundamentally the problem being solved by companies is "how do we stop journalists from agitating about our platform", not anything more interesting or important, and the "debiasing" solutions put in place reflect this reality.

Also I disagree that a "de-biased" crime model would discriminate against white men! Men commit a highly disproportionate amount of crime compared to women; any sort of adjustment you make has to adjust for that, adding a whole bunch of likelihood on women especially, probably more than the racial difference even.

You are missing the point. In de-biasing, blacks will receive an adjustment that favors them, whites will not. Women may receive some adjustment that favors them, men will not. If some model rates men negatively, this is because of the deficiencies of men. There is no need to debias the model: men are simply worse, as the model captures. If the same model rates blacks negatively, this is a flaw of the model and it must be de-biased.

This double standard is very obviously the consequence of radical anti-racist ideology. Bias is privilege + power. You can't be biased against whites or men. It is by definition impossible.

You are missing the point.

I don't think I am. I agree that a naïvely de-biased crime model will favour blacks over whites compared to a model that just went for simple accuracy and nothing else, but men will also necessarily similarly have to be favoured. If not, people are immediately going to notice the model convicting men and freeing women even when the facts are identical. There is absolutely no way people are going to accept that; radical anti-racist ideology isn't that powerful. Adding even more weight in favour of women would just be silly.

(What is slightly more realistic is if the model somehow gets access to a variable that correlates with gender but also crime itself, like your level of testosterone. With that, apologists may explain that the model convicted a man for e.g. murder based on his hormone levels which made it likely that he'd been aggressive; when in reality the model considered that to be rather unimportant compared to it being able to figure out that it's analysing a male.)

I don't think I am. I agree that a naïvely de-biased crime model will favour blacks over whites compared to a model that just went for simple accuracy and nothing else, but men will also necessarily similarly have to be favoured. If not, people are immediately going to notice the model convicting men and freeing women even when the facts are identical. There is absolutely no way people are going to accept that; radical anti-racist ideology isn't that powerful.

People have already noticed this IRL and people already accept it just fine, no radical anti-racist ideology needed. It's just the reality of the situation, sans any sort of ideology, that this sort of bias is fully and openly accepted.

But to the actual point of the thread, I think you are missing the point. I don't think sulla is describing a crime model that's de-biased "naively," but rather one that's de-biased in the most likely way that it is to be de-biased, which is by explicitly putting the thumb on the scale against disfavored groups such as whites and men. A universe in which real de-biasing efforts implemented by real institutions tend to follow some "naive" implementation rather than a politically convenient one seems like a neat universe to live in, I imagine.

People have already noticed this IRL and people already accept it just fine, no radical anti-racist ideology needed. It's just the reality of the situation, sans any sort of ideology, that this sort of bias is fully and openly accepted.

Yes, but I could probably have been more clear: I am not claiming that society will demand AI models that necessarily treat men more fairly than we do today! A model with no anti-bias applied will consider men by by default to be extremely likely offenders, especially for violent crime. It is likely that any model can get a good training score by just looking at the gender and ethnicity, and if it's e.g. an Asian woman just let her off the hook immediately.

This effect will be sufficiently extreme to get noticed, and counteracted, by adding bias in favour of men or against those women – likely not enough to make the model as a whole to favour men more than women, but it will still be adjusted away from reality in a way that favours men! An AI that randomly decides to imprison men 50% of the time and women 10% of the time can still be biased against women if women commit 0.1% of the actual crime.

In sulla's initial reply he stated that the model will be biased in favour of blacks, and biased in favour of women, which are both true but only true if you use two different definitions: "manually adjusted to favour a group" or "returning different results for different groups, all else being equal". I assume people think my reply denied that women will be a favoured group under the second definition; I do not.

I don't think sulla is describing a crime model that's de-biased "naively," but rather one that's de-biased in the most likely way that it is to be de-biased, which is by explicitly putting the thumb on the scale

That's precisely what I meant with "naïvely", as opposed to other complicated schemes (such as the case with generative AI where you could potentially do tricks like adding "no discrimination" to the prompt or the like). Apologies if that was unclear.

A model with no anti-bias applied will consider men by by default to be extremely likely offenders, especially for violent crime. It is likely that any model can get a good training score by just looking at the gender and ethnicity, and if it's e.g. an Asian woman just let her off the hook immediately.

This effect will be sufficiently extreme to get noticed, and counteracted, by adding bias in favour of men or against those women

And this is where I think you're missing the point. Perhaps the effect will be sufficiently extreme to get noticed; extreme discrepancies refuse to get noticed all the time depending on political expediency, but this could be one of those that does get noticed. It doesn't follow that there will be any desire to counteract this by people who tend to push for de-biasing such algorithms for the purpose of demographic-based justice.

Also, I don't know how to square your above explanation with:

I am not claiming that society will demand AI models that necessarily treat men more fairly than we do today!

Today, people notice that human-based sentencing systems are "biased" against men in the sense that men and women with equivalent records and crimes get sentenced very differently, with men getting more harsh sentences. People evidently have no issue with this apparent "bias" regardless of whatever lip-service they might pay.

You seem to be claiming that people will notice that AI-driven sentencing systems are "biased" against men in the sense that men and women with equivalent records and crimes get sentenced very differently, with men getting more harsh sentences, and that people noticing this will want to counteract that "bias" by putting their thumb on the scale in favor of men in these AI models. This seems to me to be similar to society demanding that AI models treat men more "fairly" than we do today.

There is absolutely no way people are going to accept that; radical anti-racist ideology isn't that powerful

Yes they will, and yes it is? We are already passed the point of naked favoritism (look at SAT scores required to get to top universities, segmented by group). There are some complaints, but most (of those who count) are happy to accept it.

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