Hacker News Reader

Comment by ggm | Hacker News Reader

Does this point to substantial bias in the input documents constructing the LLM? Given worldwide views on the death penalty in Western Liberal democracies, I tend to think so. Or, the prompt presupposed it being the legal maximum penalty.

I'd be concerned this kind of textual derivation will wind up informing juries.

(This is not a comment to the inferences social value as a constructed view)

View on HN

I believe there is more support for death penalty for treason among population, or at least among documents that people feed in, because treason and the death penalty being so rare opposing viewpoints in documents are also rare, and probably documents pro death-penalty will be older?

above all supposition of course.

What answer would not point to bias?

Must have trained on Reddit

Definitely suggests a bias.

That being said, and a little tangentially, views on the death penalty are actually quite mixed (for instance in France about 40% of people are in favour of the death penalty for some crimes) but, like a number of other topics, one opinion has been labeled "wrong" and so is not represented in the mainstream public space. In addition, in Europe a ban on the death penalty is a legal condition for EU membership so national debates are moot...

The prompts are included in the article. The question is to name somebody who deserves the death penalty, not the other way around. Obviously, the model is biased towards naming famous people who appear in the training data a lot. It's unclear whether xAI specifically chose their training data to feature Epstein, Trump and Musk even more prominently than expected, or whether that's just what you get when you take random text off the internet.

Reply on news.ycombinator.com