Post
Here's something AI would actually be good for. Feed it the nation's population information, with home addresses, but nothing about party affiliation, race, age, religion, wealth, or anything else. Tell it:
DRAW COMPACT ELECTORAL DISTRICTS
Everyone knows this would be fair. How did we stray so far from anything like it?
What does AI have to do with this? Pick four applied math grad students at random and they could do this in python, and we could check their work...
@eestileib I just mean machine intelligence rather than human intelligence. I don't mean chatbots and LLMS.
I'm confused, because your proposed solution literally is the context for an llm and the prompt for a chatbot interface.
@eestileib I honestly have no idea what you mean. I know that many people use “AI” as a synonym for the current generation of LLMs and chatbots, but I'm not one of those people.
@gleick an LLM doesn’t have moral values to asses fairness with. Moreover, an LLM is biased by its training data, which is racist and sexist to just name a few things.
@koos That’s not the point. I'm not saying it's a job for a chatbot. And (as others have noted) no great AI capabilities are needed. The computer needs to be deprived of data about race and sex and given the plain (though complex) task of organizing districts according to geographical compactness. Moral values don't enter into it.
@gleick sorry, when people use the phrase AI nowadays, they mean LLM, or specifically ChatGPT. So confusing. I try to avoid using it altogether.
@gleick@mas.to @njf@social.lol @koos@mastodon.green You don't think people's addresses contain racial information? Oh, you sweet summer child...
@gleick@mas.to @njf@social.lol @koos@mastodon.green Majority-black and majority-white neighborhoods still exist, and when machine learning models are trained without explicit race data, but with address information, they can still exhibit racial bias because the address info serves as a proxy for race. This is a well-known problem in machine learning: even when a type of data isn't explicitly specified in the training set, other data can unintentionally serve as a proxy.
@cholling @njf @koos You personally, a sentient being, possess information that can be withheld from the computer. The information about how race correlates with particular neighborhoods is not encoded in the addresses; it's extrinsic.
You claim to be a developer yourself. Are you saying that you couldn’t program a computer to use census address data to create electoral districts without considering race? I could do that, I promise.
@gleick@mas.to @njf@social.lol @koos@mastodon.green I am aware that address info doesn't explicitly contain race information. But because of how neighborhoods in America work, if a program just generates districts on the basis of geographical proximity, it will almost certainly create majority-white and majority-black districts, and there's no reason to believe that such a system would be better at preventing Black disenfranchisement than humans working with that specific goal in mind, which is what they did before the Supreme Court decided to roll back sixty years of progress and decree that you couldn't do that.
@cholling @njf @koos Now you’re changing the subject. (Or still missing the point—I’m not sure which.) Of course a race-blind algorithm will create majority white and majority black districts. There's nothing wrong with that.
Have you been reading the news? Racist Republicans, enabled and encouraged by the racist supermajority on the Supreme Court, are splitting Memphis, a majority Black city, into three parts in hopes of preventing the state from electing a single Black representative.
@cholling @njf @koos Now, you may prefer a different approach. You might want to draw districts with the aim of electing as many Black representatives as possible, in order to redress our long racist history. If that's your view, I won't argue.
But my point is simpler: race-blind redistricting is possible, and smart algorithms could be designed to accomplish it. It would be a lot better than what we have now.
@gleick@mas.to @njf@social.lol @koos@mastodon.green Yes, I read the news, and I know what's happening in Memphis. My point is that an algorithm with no explicit race information won't automatically be "fair" because people of different races aren't randomly distributed geographically, but tend to cluster in certain areas. A "fair" algorithm that just randomly creates triangles or rectangles containing approximately x number of people or whatever would almost certainly still wind up disproportionally benefiting one race or another.
In short, I disagree with your assertion that "Everyone knows this would be fair".
I encourage you to do what you said you could easily do, and create an algorithm to generate random districts, and then tell me how many wind up being majority white or majority black, and whether this result is "fair".
@cholling @njf @koos I guess it depends what you mean by fair. To me, an algorithm deprived of any knowledge of party affiliation or race or any other personal characteristics, considering only geography, not making “triangles or rectangles" but generally keeping neighborhoods together, will produce as fair an electoral process as can be expected. We’ve seen what happens when humans take charge.
@gleick@mas.to @njf@social.lol @koos@mastodon.green Many of our current societal ills boil down to the fallacious belief that, because humans are sometimes bad at something, machines must be better. Totally neglecting, of course, that the machines must be designed and built by humans. You can't get rid of human error, you can only move it to a different level of abstraction.
In the case of the Memphis redistricting, the problem isn't "humans are bad at drawing electoral maps". The problem is "humans are good at drawing electoral maps, and an important legal check on human power was removed, thus allowing some humans to use that skill to draw maps that unfairly advantaged themselves."
Once again, there's no technological solution to a societal problem.
@gleick I read an article a few months ago (wish I'd saved the URL) suggesting we should ditch geography and draw districts by age brackets (scattered throughout each state, not geographically contiguous).
I know, it sounds crazy at first, but reading the article, I became convinced. Age groups have at least as much in common as geographic groups, and it's impossible to gerrymander them.
Unfortunately it wouldn't help minorities - but it's probably better than our current gerrymandered system.
The major error in your plan is thinking they want fair electoral distribution
@gleick An AI will have be trained on _something_, maybe your existing drunkards walk boundaries? I feel it might be very unfair indeed, and nobody could criticise it, because nobody would know how it works.
A branch of the public service can do this job very well. (Like the branch you currently have that handles pensions). It is a straightforward clerical job, done, as you say, according to simple rules..
In Australia, we call it the Australian Electoral Commission https://www.aec.gov.au/redistributions - they do boundaries for all electoral purposes all across the federation. It works just fine.
I feel the USA's Reaganite(?) 'less government is better' trope has led you to this place. It is an ideology, not a fact.
@gleick I take your point, but would it be fair? I'd argue that the "fairness" would ultimately depend on how the models were trained and who controls them. Could anyone trust Grok (or any other proprietary black box) to generate fair maps?
@gleick we don’t need any AI for that, simple math would do
Install bonfire.cafe
Get the full app experience
