Hacker News Reader

Comment by camillomiller | Hacker News Reader

I was enjoying this until this signature sentence by Claude:

>>Slowly but surely, it did, once again becoming a world-renowned brand, now separate from that of Thonet. The same town, the same factory, the same people—just a different name

The moment I see this, the piece loses all of its credibility. I cannot take any of the factual informations I was enjoying as implicitly correct now, I just can’t. Authors should edit LLM help in a much stricter way. As a writer, LLMs should only be research helpers and used to check for typos and syntax mistakes. Authors, please, do better

View on HN

don’t be too quick to judge, putting the text through pangram indicates it’s 100% human written. It didn’t strike me as too AI because it wasn’t superfluous; it literally was what happened. AI is using this phrase more as a stylistic device than an a factual statement

"I Wouldn't Say Pangram is Broken, But I Would Say That It's Brittle": https://freddiedeboer.substack.com/p/i-wouldnt-say-pangram-i...

> it wasn’t superfluous; it literally was what happened.

Yeah, so I gave that sentence to ChatGPT in a new context and asked it to make it longer without saying anything additional, several times, and it still reads fine, more than. The truth is the truth no matter how you put it.

https://pastebin.com/gKhCbmNQ

https://pastebin.com/9HpkN4tc

https://pastebin.com/J6znwxZf

https://pastebin.com/wXs84csy

https://pastebin.com/LrjVm4JT

The last one really speaks to me. I wanted to go on but it raised all sorts of questions in me that I think I should investigate instead.

It was, in that sense, a change that changed the thing that was being called something without changing the thing that was being called.

and

Everything that was not the name remained the same.

Everything that was the name was no longer the same.

Really makes you think.

Sometimes I'm taken aback by the visceral reaction some people have when their spider sense detects LLM impurities in the writing.

I read that same sentence (and the whole article, which I loved) and it didn't trigger anything in me.

It reminds me a bit of the OMG GMOs! craze people used to have a while back (and might still have?).

I think it's more likely model creators will up their game. I've been generating some copy for our company website. I'll be honest, not all of it is great and there are clear AI slop signs. However, you can use guard rails to limit this kind of thing somewhat. I've been doing that. But the good news is that there are pretty easy to spot patterns. And LLMs are actually really good at identifying those already. It seems spotting the patterns is a lot easier for LLMs than addressing them.

> As a writer, LLMs should only be research helpers and used to check for typos and syntax mistakes. Authors, please, do better

I know professional writers that are definitely no longer just limiting themselves like that. I think that ship has sailed.

Everyone with a stake in AI wants that ship to have sailed.

You do know em dashes are a normal part of formal prose right? If you go through pre GPT New Yorker and the Atlantic articles, there are plenty of em dashes too.

The New Yorker has a whole style guide. Most people aren't writing "née" or "reëlected" very often but the New Yorker keeps those formats and even has a custom typeface.

Yes em-dashes are real but aside from a few NY-based magazines with real editors, writers aren't actually using them very much. I'm a bit sad because I like using them and now I think twice before doing so in case it is misinterpreted as an AI signature.

Also people are generally more aware of the American-English usage of them. The UK (and splitters) uses them differently, as do non-English languages.

It has nothing to do with the em dash, you AI-dweeb. The entire construction is pure LLM slop. Go touch some AI-generated grass.

I think GP was talking about the phrasing itself, not the dash. I actually didn't notice the em dash itself when reading it and agreeing.

Reply on news.ycombinator.com