Hacker News Reader

Comment by psadauskas | Hacker News Reader

I've mostly stopped caring about using using proper capitalization, commas, grammar and spelling in my writing of comments, primarily as a signal that i'm not an llm.

View on HN

If you turn on HN's "Show Dead" setting, there are tons of LLM-generated comments on stories related to AI. You can see the human(s) behind the LLM trying to fiddle with the style of comment by making them skip proper grammar, capitalization, use or avoid certain phrases, and so on. The biggest tell for LLM content, though, is just the content as a whole: it sounds fake and ungenuine, like it passed through a committee of hostage negotiators to remove the speaker's own attachment/expectations.

They can configure it to use all lowercase letters, skip em-dashes, make grammar mistakes, stop saying "it's not X, it's Y", or whatever, yet the content itself just has a fake quality to it that makes it stand out, which is why those comments still get flagged IMO.

The uncanny valley of text. It looks and sounds like a human, but lacks the "soul" / humanity that our intuition somehow perceives.

It's really strange... I see some text with obvious tropes and sometimes I read something and there's no obvious AI trope... but it's just not human?

> The uncanny valley of text.

Exactly, that's a great way to describe it.

And yet, while on HN we're critical readers and can still see through it, there's many places on the internet where it just wouldn't stand out. I try to avoid them, but they would just blend in to e.g. youtube comments.

Unless the YT comments I've read have been bots since forever.

My big gripe with LLMs is the “high verbosity but low signal” style of their writing. Even the new 4.8 Opus writes like that, it’ll spit out a paragraph of verbiage for what could’ve been one sentence. I’m guessing… because we pay by output tokens? $-)

i mean, i definitely agree and am somehow allergic to seeing llm written text from other humans (you typed a prompt! why not just post it directly? i'd rather have bad spelling and grammar than llm slop). but... while i feel i can detect it pretty reliably in forums like HN, i can't help but think that this is the toupée fallacy[0] at work. of course, all the text that _i_ think is fake is clearly fake after all

[0] https://rationalwiki.org/wiki/Toupee_fallacy

I had this conversation the other day. I'm a native German speaker originally, which is why I hand out commas like it's candy and capitalize things unnecessarily. Sometimes I notice these things and leave them in when I write something, since at least it gives you a good indication that a human wrote it... for now.

I stopped doing that long before LLMs were commonplace because I couldn't see a point in it. Like, the entire concept of spelling and grammar is arbitrary anyway. Proper English and spelling of the 20th century is not the same proper English and spelling of the 18th century.

For example, "you" was originally the formal form of the second person pronoun, and thee or thou the informal form. Many writers who try to write midieval period pieces tend to get this wrong though and just use thee or thou as a direct replacement for "you."

And then English spelling and pronunciation is just chaotic anyway.

I won't go out of my way to misspell things and I'll do my best to use the best grammar and spelling I can, but I'm not going to consult an llm or grammarly to make sure it'll get no notes from an English teacher when my only purpose is to comment on HN or write a quick update on slack.

Claude's "write me a product description like a cool human would" is just using lower-case where it shouldn't be though.

The problem is that omitting capitalization, commas, and so on signals, in addition to "not AI in default settings", but also "I'm part of the San Francisco AI in-crowd and Altman is my spirit animal".

Reply on news.ycombinator.com