Hacker News Reader

Comment by ivanbakel | Hacker News Reader

I think you're missing the nuance of the author. "Bullshit" is not synonymous with "wrong information", it means something that is produced while disregarding the truth.

AIs don't "sometimes produce bullshit", they are always producing bullshit. Sometimes what they produce is accurate, as a consequence of the initial corpus being accurate. But to the AI, much like to the demagogue, this is totally irrelevant - it only has to sound good.

And if it's "easy" to verify what AIs say, then it should be "easy" to teach AI to verify before it speaks, which would automatically counter this criticism of AI as bullshit machines. But the real threat is that AI can produce bullshit which is hard to verify, and can do so at a speed that puts professional gishgallopers to shame.

View on HN

> And if it's "easy" to verify what AIs say, then it should be "easy" to teach AI to verify before it speaks

Sure, if that's the only thing you want the LLM model to be able to do, you could do that. But there are other use cases beyond getting facts, that these models usually also want to support.

> it means something that is produced while disregarding the truth.

The Cambridge Dictionary says it has two definitions[1]:

1. a rude word for complete nonsense or something that is not true

2. a rude word meaning to try to persuade someone or make them admire you by saying things that are not true

I don't think thise align with your definition and your idea that AIs always produce bullshit. They don't always produce nonsense or something that isn't true, and they don't intentionally use untrue statements to try and persuade someone or gain admiration.

[1] https://dictionary.cambridge.org/dictionary/english/bullshit

The definition of bullshit used here is not the dictionary one. It more closely follows the definition from On Bullshit[0] (also cited in TFA), as content which is truth-irrelevant.

Under the above definition, AIs do always produce bullshit, because their content is always truth-irrelevant.

[0]: https://en.wikipedia.org/wiki/On_Bullshit

Following the pre-eminent scholar of bullshit, Harry Frankfurt, we can attempt a ‘bullshit alignment matrix’ according to the ground truth of a statement, and its intent. - https://www2.csudh.edu/ccauthen/576f12/frankfurt__harry_-_on...

On one axis we have the truth value of a statement, on the other, the intent (sadly no tables on HN)

Good intent

- Factually truthful: 'Roses are red'

- No inherent truth Value: Maybe surreal humor, like 'this statement is humor'

- Factually false: Satire, parody, sarcasm, 'Great HN comment Einstein'

Bullshit: Statements for effect, indifferent to truth value

- True: 'More doctors smoke Camels'; 'My response to the Covid pandemic was amazing!';

- No truth value - ill-formed nonsense, oxymoron, non sequitur or empty puffery: 'Our Founding Fathers under divine guidance created a new beginning for mankind.'; 'Starbucks provides an immersive ultra-premium coffee-forward experience.'; 'Nobody’s bigger or better at the military than I am.'

- False: 'Obama was born in Kenya'; 'We had the biggest inauguration crowd ever.'

Bad intent, to defraud or mislead:

- True: 'I did not have sexual relations with that woman, Miss Lewinsky'

- No truth value: 'We’re not going to sit here and listen to you bad-mouth the United States of America!'

- False: 'Clinically proven to boost genes and make your skin visibly younger in just a week'; Gish Gallop

Of course this is mostly bullshit:

- Hard to discern intent when something turns out to be false, was Bush's WMD claim bullshit, or a lie; similarly Obama, 'if you like your insurance, you can keep it'. One man's benign public health simplification about masks or food pyramids is another's nefarious conspiracy.

- Even hard science isn't free of bullshit, Feynman's cargo cult speech notwithstanding, any sufficiently complex science and engineering devolves to some amount of cargo cult bullshit in practice

- Let's not even get started on organized religion, Washington and apple trees and whatnot. Communication takes place on multiple levels, fruit flies like a banana, something can be bullshit on one level and good (or bad) intent on another.

In the words of George Carlin, "Bullshit is the glue that binds us as a nation." Or Napoleon, "History is a set of lies agreed upon."

GPT's ability to bullshit at a human level is a truly monumental achievement.

'this statement is false' is what I mean to write, not 'this statement is humor' , self contradictory or ill-formed nonsense that has some meta value like humor

I wrote a long-form version of this bullshit here - https://druce.ai/2023/09/bullshit

That's a bit too much hyperbole for me. Claiming that something always produces bullshit gives the impression that it is useless, which LLMs are clearly not.

Reply on news.ycombinator.com