Words Make It Obvious That Your Text Is Written by AI | Hacker News Reader

Words Make It Obvious That Your Text Is Written by AI

medium.com

14 pointsdoener8 comments

My ad-hoc radar for GPT generated text is:

  1. Is the writer a 5th grade student?
  No --> GOTO 2
  2. Is the text structured like an MLA 5-paragraph essay with prosaic and insipid content?
  Yes --> GOTO 3
  3. Does the final paragraph contain some variant of the phrase "in conclusion"?
  Yes --> High probability of being LLM generated

Could it be the training data contained a large trove of generic essays used by college students to cheat on writing assignments?

I mean, maybe it's already been suggested, but I haven't seen anyone come right out and say it.

Your wording is interesting--I don't think you're wrong, but mentally rephrased as "a large trove of essays in the generic format mandated by teachers".

It seems tragic that so many students were/are forced into this mechanical turk without pay, for work that is meant to be discarded, in a format that is flatly unacceptable, and now suspicious, in any other context.

I don't really agree with this article. I think that those particular points are actually just commonly used in typed dialog....because they are common in spoken casual explanation.

those particular examples are all 'descriptive' or educational to me, and alternatives were not suggested with any concrete examples.

They are fairly common, but there is something to it. I've recognized a lot of these turns of phrases unconsciously, noting that it sounds like AI, but not being able to put my finger on why.

It's like ChatGPT uses these more uncommon terms and phrases somewhat more often than the average person. And it'll often give multiple of these tells in the same conversation.

... sounds like something an LLM would say!

With more of these articles surfacing all prompts will just have a built in: "Don't include these words: Don't use any of these phrases: Etc..."

Honestly, I don't even find this article very well-written. I wouldn't be surprised if it was at least partially AI-generated.

We definitely need ways to verify if content was properly human-sourced, but this ain't it.

Reply on news.ycombinator.com