Hacker News Reader

Comment by Imnimo | Hacker News Reader

There are a LOT of fundamental misunderstandings underlying this experiment, but I think the biggest is this:

Why does the idea of a "political quiz" work for humans? Because humans are at least somewhat consistent - their answer to a simple question like "would you rather have a smaller government or a bigger government?" is pretty informative about their larger system of beliefs. You can extrapolate from the quiz answers to an ideology, because humans have an underlying worldview that informs their answers to the quiz. ChatGPT doesn't have that. If it says "bigger government", that doesn't mean that its free-form outputs to other prompts will display that same preference.

Trying to ascribe a larger meaning to ChatGPT's answers to a political quiz tells me that the author is very confused about what ChatGPT is and how it works. It's too bad Reason doesn't have an editor or something who can filter this sort of nonsense.

View on HN

In the 90% case, people's "wants" are absolutely not "consistent" for any definition of consistent. Inconsistent/contradictory answers can be easily elided through question framing. Ask someone who responds to a "smaller vs bigger government" question with "smaller" something about concrete services that may or may not be provided by governments, and in most cases you can elicit support for much larger government apparatus than exists now anywhere.

All people are shaped by unique experiences and productions of those experiences can be "consistent" and can hold firm in the face of even contradictory evidence.

But in the absence of deep personal experience, people, like ChatGPT, are argument-producing machines, where the shape of the produced arguments depend more on prompt, social context like the status of other parties in the argument, and personal context like whether someone is hungry or tired, etc.

Thank goodness the for-centuries-incorrect models of people as primarily intellectual creatures are subsiding and the reality of people being storytelling creatures are coming to the fore.

Relatedly, agree completely that Reason lacks informed editing on many topics. Cheers.

It is true that ChatGPT's answers do not have a deeper underlying meaning because it does not have a consistent model of the world. Many humans don't either. However, it is possible for a chat model to be biased in one way or another. This can happen through human judgment and rating of its answers (as mentioned in the article), or through the selection of the materials it is trained on. So, it is not unreasonable to try to understand its biases because it will be used by students.

I strongly agree. ChatGPT ultimately tries to give you what you ask for - so if certain prompts cause it to give "leftist" (by whatever definition) answers, then one might reasonably conclude that something about those prompts makes ChatGPT think you wanted a leftist answer. But that's wildly different from claiming that ChatGPT itself has a leftist bias - which is a bit absurd considering that it will happily give out virulently anti-leftist answers, if that's what it thinks you asked for.

ChatGPT is the ultimate example of GIGO - Garbage In Garbage Out.

Although you could argue some of the non-answers to questions concerning African American diaspora, porn, eugenics etc suggests that the overriding of the AIs responses by the humans running OpenAI has lent it a left-leaning bias.

What are some of these non-answers? Dare I ask at all?

> Why does the idea of a "political quiz" work for humans?

I'd argue it doesn't work for humans neither. It makes you barely scratch the surface of politics and then you double down on the results without understanding any of the arguments made in the past 200 or so years that lead to the current situation.

It also doesn't help that every quiz I took is just bad. Oh, you're not opposed to same-sex marriage? Boom, lib left. About as anarchist as Emma Goldman who was against prosecuting gay people in the beginning of 20th century (unheard of at the time).

> It's too bad Reason doesn't have an editor or something who can filter this sort of nonsense.

Editors of newspapers are supposed to know how AI models work?

> Editors of newspapers are supposed to know how AI models work?

Editors of newsmagazines are supposed to know how journalism works, and to probe journalists to assure that they’ve done the research to properly understand and represent the subjects of their stories. So, yes, when publishing a story about AI models, the editor should know enough about how AI models work to understand that the story is a proper representation, not in advance of seeing the story, but as a result of the dialogue with the reporter on the research that the reporter has done before the editor signs off on the story.

Parent did not say they should understand ML on a fundamental level. Reading the description in the comment should be enough for the author of the article to say "oh" and not write this article.

...because the editor understood he misunderstood how the algorithm works? Or becuase they don't know so they shouldn't risk signing off on it?

Sorry I don't see how you could possible challenge this article without a deeper understanding of the technology. That is the fundamental critique here is it not?

Expecting every newspaper to have those domain experts on hand for each topic is naive and highly optimistic. The journalism world doesn't work that way for better or worse.

If it did we would have 50%+ articles never written. Probably more.

> If it did we would have 50%+ articles never written. Probably more.

Doesn't sound like a very bad outcome honestly.

"The flip side of that is that it closes the valves of one's attention"

- from the movie Tar

the point is ChatGPT seems too consistent

I don’t know how well such quizzes align with ideology. You have hard core republicans that love Trump and hate the left, etc, but also want a nanny state to protect them from the “false advertising” Of McDonald’s good pictures, and hate capitalism abs concept of “supply and demand”. Then you have enormous amount of relatively well of leftists that claim they want they want a socialist egalitarian utopia and clearly know full well they don’t want the consequences of such a society to fall to them.

Reply on news.ycombinator.com