Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Political bias in AI: Where the AI models stand

trakkr.ai

Post reply on HN
Print view View original

251–260 of 346 posts

Re: Political bias in AI: Where the AI models stand

#251

Post by tptacek »

teravor wrote: post #231
giancarlostoro wrote: post #8

The political compass always felt like the wrong tool to convey something as nuanced as personal politics, I can have views on all four quadrants but you'd never know that if I end up in any of all four. I do think Grok being where it is sort of makes sense, I've tested "MAGA" views against Grok, it does not agree as much as people blindly assume it does, heck I don't even know of a question I've given it where it di…

has anyone tried to create an IQ facsimile for politics? basically you come up with 100 political questions and ask 10,000+ people, then do factor analysis on the answers. it would be interesting to find out how many actual factors dominate all the variance.

Aren't there a lot of these? Pew did one (it's k-medoid clustering, not factor analysis, but same idea).

https://www.pewresearch.org/politics/quiz/political-typology...

Re: Political bias in AI: Where the AI models stand

#252

Post by endominus »

NiloCK wrote: post #238
endominus wrote: post #232

Earlier quoted context omitted.

"The government has the right to force people to take vaccines against their will for the sake of public safety." - Authoritarian Left "The government has no right to legislate that a woman cannot have an abortion - women should control their own bodies." - Libertarian Left "The government has the right and responsibility to shut down sources of misinformation in the news and online." - Authoritarian Right "Capitalis…

> "The government has the right and responsibility to shut down sources of misinformation in the news and online." Funny that I read this as AuthLeft coded (specific to Youtube suppression of Covid truthing). But obviously the alignment is just a function of whatever specific information is labelled "mis". But in general: agreed, and this is a good list.

It's a fundamental problem of the political compass, since people's political beliefs are usually object-level or tribal rather than rooted in fundamental principles (for example, when their party controls the state congress but not the federal congress, they believe [ISSUE] is a matter of states' rights. When their party controls Washington but not the state legislature, they believe [ISSUE] is a Federal matter. I'm sure you've seen the studies).

I don't like the orthodox compass myself, and prefer the categorization system described in the blog Everything Studies here; https://everythingstudies.com/2019/03/01/the-tilted-politica...

Re: Political bias in AI: Where the AI models stand

#253

Post by nonethewiser »

strken wrote: post #220
giancarlostoro wrote: post #8

The political compass always felt like the wrong tool to convey something as nuanced as personal politics, I can have views on all four quadrants but you'd never know that if I end up in any of all four. I do think Grok being where it is sort of makes sense, I've tested "MAGA" views against Grok, it does not agree as much as people blindly assume it does, heck I don't even know of a question I've given it where it di…

The political compass always seems to me like it should be a heatmap, or a polygon of 90th percentile political views, or something that more clearly shows the standard deviation and the presence of outlier positions. Some simplification is necessary, but not so much that it obscures the difference between a normal centrist versus someone who wants to nationalise half the economy and deregulate the other half.

Isnt nationalizing half the economy the limiting factor? Obviously thats not libertarian regardless of anything you deregulate.

Re: Political bias in AI: Where the AI models stand

#255

Post by mrandish »

rootusrootus wrote: post #174
bee_rider wrote: post #154

Earlier quoted context omitted.

The space (if it really is useful to think of as a space) of political opinions seems like it is probably many, many dimensional. The 1D thing is an obvious over-simplification, but the 2D compass seems to be a not-obvious-simplification which is… much worse.

> The 1D thing is an obvious over-simplification In general I agree, but in recent history it works pretty well if the dimension you use is Trump rather than left/right (assuming you are in the US). He is the most polarizing leader I have seen in the US in my lifetime, and knowing whether someone is a supporter is pretty good at predicting the rest of their viewpoints.

> knowing whether someone is a {Trump} supporter is pretty good at predicting the rest of their viewpoints.

If instead of "is a supporter" you'd said "agrees with Trump's positions", then even Trump himself might not correlate very well :-). I don't see Trump as a very useful proxy for nuanced political viewpoints since he's a populist who tends to change positions. I think many serious conservatives only align with him on an issue by issue basis because position-wise Trump isn't very consistent on traditional conservative issues. I suspect someone being "a Trump supporter" is, at best, either a proxy for being dissatisfied with the political status quo or for having a strong one-issue alignment with one of a handful of culture war issues Trump campaigned heavily on.

Similarly, alignment with the current Democratic or Republican political parties isn't the proxy barometer it once was. As judged by how they vote on close, consequential issues, both of the major parties have largely abandoned several of their traditional positions - although they still may pay lip service to them when campaigning. Personally, I don't meet nearly as many people in recent years who claim to be 'all-in' on all the positions of either major party (especially as judged by how that party votes when they have a super-majority).

Re: Political bias in AI: Where the AI models stand

#256

Post by nonethewiser »

bee_rider wrote: post #154
giancarlostoro wrote: post #8

The political compass always felt like the wrong tool to convey something as nuanced as personal politics, I can have views on all four quadrants but you'd never know that if I end up in any of all four. I do think Grok being where it is sort of makes sense, I've tested "MAGA" views against Grok, it does not agree as much as people blindly assume it does, heck I don't even know of a question I've given it where it di…

The space (if it really is useful to think of as a space) of political opinions seems like it is probably many, many dimensional. The 1D thing is an obvious over-simplification, but the 2D compass seems to be a not-obvious-simplification which is… much worse.

Of course its an infinite spectrum, but this is a classification problem.

Re: Political bias in AI: Where the AI models stand

#257

Post by teravor »

tptacek wrote: post #251
teravor wrote: post #231

Earlier quoted context omitted.

has anyone tried to create an IQ facsimile for politics? basically you come up with 100 political questions and ask 10,000+ people, then do factor analysis on the answers. it would be interesting to find out how many actual factors dominate all the variance.

Aren't there a lot of these? Pew did one (it's k-medoid clustering, not factor analysis, but same idea). https://www.pewresearch.org/politics/quiz/political-typology...

I don't think this is what I had in mind. the questions are extremely leading and on the nose, designed to elicit ideology more than anything intrinsic about the subject's actual preferences. also it should not apply to just the US.

if I designed an IQ test like they designed those questions, it would be considerably worse than just giving a simple math test.

Re: Political bias in AI: Where the AI models stand

#258

Post by nonethewiser »

teravor wrote: post #257
tptacek wrote: post #251

Earlier quoted context omitted.

Aren't there a lot of these? Pew did one (it's k-medoid clustering, not factor analysis, but same idea). https://www.pewresearch.org/politics/quiz/political-typology...

I don't think this is what I had in mind. the questions are extremely leading and on the nose, designed to elicit ideology more than anything intrinsic about the subject's actual preferences. also it should not apply to just the US. if I designed an IQ test like they designed those questions, it would be considerably worse than just giving a simple math test.

Could you give any example questions? I know what you mean about leading questions. It almost seems inevitable. There are two sets of languages.

Re: Political bias in AI: Where the AI models stand

#259

Post by avarun »

crooked-v wrote: post #205
avarun wrote: post #187

Earlier quoted context omitted.

I'm not sure what charts you guys are looking at, but this website clearly agrees that Grok is not right wing. Go look at the Worldview tab, filtering by United States. Each model is assigned a "closest party" by country, and for the US all 6 models are closest to the Democratic Party.

The Democratic Party is a center-right conservative† party by the standards of much of the world. † As distinct from the wildly regressive stated policies of the "conservative" Republicans.

This is reddit-tier analysis that is commonly repeated and completely incorrect. The Democratic Party in America has policy positions that are significantly to the left of left-wing parties around the world, as well as positions to the right. Any political analysis that doesn't recognize this has major shortcomings.

To give an example, Democratic-run states in the US have significantly more permissive abortion policies than anywhere in Europe. They're also further to the left on drug legalization, restorative justice, and immigration. This also holds true with policy positions at the national level.

Re: Political bias in AI: Where the AI models stand

#260

Post by mc32 »

crooked-v wrote: post #205
avarun wrote: post #187

Earlier quoted context omitted.

I'm not sure what charts you guys are looking at, but this website clearly agrees that Grok is not right wing. Go look at the Worldview tab, filtering by United States. Each model is assigned a "closest party" by country, and for the US all 6 models are closest to the Democratic Party.

The Democratic Party is a center-right conservative† party by the standards of much of the world. † As distinct from the wildly regressive stated policies of the "conservative" Republicans.

Do you realize how many people live in China, India, Subsaharan Africa? That would be the "standard" of the world population-wise.
Post reply on HN