This article points out some of the common tells of AI writing using a lot of examples gathered from attempted Wikipedia edits.
Have you seen a guide like this for non-wiki writing? I think my boss uses AI for her emails. She sucks. Friend pointed out that her emails often have long hyphens which are supposedly a chatgpt thing.
OP’s article is applicable.
ChatGPT uses the same style for emails. Flowery prose, bold/italics, and huge blocks of text that don’t say much of anything. Another hallmark is “slop” phrases; overused terms you eventually learn to recognize. See this:
https://old.reddit.com/r/LocalLLaMA/comments/1g6k4ci/superslop/
https://github.com/sam-paech/antislop-sampler/blob/main/slop_phrase_prob_adjustments.json
Also check out EQBench’s slop taxonomy in the writing bench: https://eqbench.com/creative_writing_longform.html
And the slop score, which is an indicator more than a detector: https://eqbench.com/slop-score.html
Fortunately, the vast majority of non-developers (like your boss) use ChatGPT or Gemini, and their style is pretty easy to sense… I call them “deep fried.”
But different models can be harder to sniff out.
Like, if I wanted to fool you, I’d probably pick GLM 4.6 with task vectors, a slop blacklist, the force it to think in Chinese (so its own English tokens don’t bias it). Then it’d be basically impossible to detect.
Thanks for the links, saving this to peruse later 😁
If it’s sent with Outlook, em-dashes are automatically created from two hyphens by default.
Wikipedia has quite stringent standards for citations and style. Just enforcing those should be enough to keep slop out.
Removed by mod
This will tank my confidence in the accuracy of Wikipedia. Is there nothing free from AI slop? JFC
Why? This is about how Wikipedia tries to prevent AI slop making it into articles.
Thanks for the clarification. If AI is allowed to or makes it onto the Wikipedia platform - if they no longer hold back. Hopefully that makes more sense
