Hacker News Reader

Comment by LanceJones | Hacker News Reader

I always wonder how people can tell. For this particular article, was it the thirty-four occurrences of em dashes with spaces on either side? Something else obvious?

View on HN

It is the em dashes and the excessive wordiness as well as a lot of "not this, but that".

Eg:

"Not dramatically. Just quietly. " -- This is filler words. Whether it's dramatic or quiet has no relevance to the point they're making.

It also loves threes: "Well-modelled, properly sourced, beautifully visualised to requirements" - again, all irrelevant. The point they're making is that it's measuring the wrong thing, not that "beautifully visual things can be incorrect".

"There’s a piece of this conversation that most leaders miss, and it’s the part I care about most" - this hook of "most people miss" it is very common in AI writing.

The tiny sentence fragments are too much for me. They trip up the flow of the text.

Also the "not this, but that" structure is overused here.

Articles of this type suggest a fun game: "LLM or Marketroid?" Because either one could have written it, and both are capable of about equivalent levels of original thought. (whoops did i just say that out loud)

Hey, the LLMs had to pick up these patterns from somewhere. Embarrassingly, before LLMs, there were plenty of humans that wrote like this.

It was the content. So many very specific claims with no source, just stuff being made up. I don't know who Brené Brown is, perhaps she specifically researches trust, but how curious that her daughter can raise a problem with trust, specifically cite two named behaviours that build trust, and then Brown just happens to have a database of trust-building behaviours to hand, that she hasn't even analysed, ready to output a teachable moment.

https://en.wikipedia.org/wiki/Brené_Brown

I am not sure if you posted a brilliant, subtle joke... or if you're demonstrating the exact behavior that you suggest is a flag of inauthenticity.

In the article, she wasn't introduced as a researcher at all, but suddenly "She went back to her research data...". This totally smells like an LLM refactor where it re-emits surface level details, but completely misses the key beats that tie ideas together across a story.

So, failure to do research and ensure a coherent mental model leads to erroneous inferences… for humans as well :)

What's your point? Yes, humans and AIs CAN make similar errors. However, continuity and context churn at this level is the sort of error that humans make rarely and AIs make constantly.

By itself it is not a sure sign of LLM output, but along with other signals it is highly suggestive.

This particular article has the tell tale opus 4.8 smell of these short sentences. I think its mainly opus 4.8

When did "X is built one marble at a time" become popular? Maybe search analytics can tell us.

Well, this isn't what I expected. (And it only goes up to 2022.) https://books.google.com/ngrams/graph?content=one+marble+at+...

This one almost feels like the AI got stuck in a perseverating loop of "He <blank> the <blank>." <repeat>

This is followed up by a sprinkling of every possible punctuative shakeup: bold, em-dash, semicolon, colon, quote, etc.

Well, sometimes there's flat-out nonsense that seems to have been written purely to back into the author's thesis:

    You cannot design an algorithm that eavesdrops on dinner conversation and dispatches someone to buy a street hot dog, because the person on the receiving end would immediately sense the machinery of it.

But usually there's also:

- Word count hovering between four and five thousand words - Dramatic/narrative section titles - "No X, no Y. Just Z"

Last but certainly not least, there's the Lists of Exactly Three Things. I counted literally thirty in this piece. Examples:

    - "...the ritual of a human voice, the small exchange about an anniversary or a first date, the warmth of being recognised."
    - "Who was celebrating a birthday? Who was on a first date? What had a regular not finished on their plate six months ago?"
    - "You can’t purchase it, automate it, or accelerate it with a clever marketing campaign."
    - "...forgive outages, laugh off a late delivery, stay through a price increase."
    - "...the food arrives hot, the bill is accurate, the room is clean."
    - "You notice, you adjust, you respond."

I personally thought to myself "written by AI" after this part:

... the restaurant was fully booked. No warmth. No conversation. Just a long wait and a closed door. In trying to humanise the process, he’d made it worse.

I'm sure some people write this way, but most don't. And AI writes this way.

Reply on news.ycombinator.com