Hacker News Reader

HNHacker News

Top New Best Ask Show Jobs

Comment by lilerjee | Hacker News Reader

Full thread

lilerjee·

How to detect which fiction, article, or review is created by fake AI or LLM?

There are too many texts generated by LLM on Internet now.

The experience of exploring Internet is very bad now.

View on HN

It's often quite easy, because most people who author articles with LLMs do not bother doing any kind of post editing. Meaning you can see/build a mental map of common tropes[^1] for what to look for. Once you read perhaps two or three articles authored with LLMs you will not be able to unsee it.

[^1]: https://tropes.fyi/

Yes but it's still like wading through shit to find the occasional gem. It's not hard to pick out when you work with LLM's a lot. But it's still a bad experience. Social media was already hard at work at killing the internet (or at least what the internet used to be). LLM's made it worse.

For fun, I pasted a few Matt Levine's newsletters into the AI Vetter on the above page. Since it can't fetch the Bloomberg's paged directly, I had to copy the content into Github gists.

4 out of 5 were marked as suspicious, mostly citing short punching sentences, negative parallelism, and content duplication.

Look at the publishing date. Unfortunately I don’t think we will have a more reliable way to do it

If the artist isn't one that I already know and trust, then this is how I do it.

It's very sad, but I have to view any recently-created work from an unfamiliar artist with great suspicion. It does make me less interested in recent works by default, because of the work of vetting the artist.

Reply on news.ycombinator.com