Found a video guide to spotting posts written by LLMs. Thought it might be useful.
TL;DR for AI writing warning signs:
- Use of the em-dash (—)
- Parallel sentence structure (e.g. "It's not just X, it's Y")
- Grouping things in threes or at least odd numbers
- Delineating line breaks with emojis
- Odd/unnatural verbiage
- Overuse of filler words (talking like your average LinkedIn post)
- Exaggerated and empty praise
- Weird analogies and similes
- Restating and overclarifying points
TL;DR for signs something was written by a human:
- Including anecdotes
- Written in the first person
- Tangents and nonlinear storytelling
View original on lemmy.world
173
Comments58
As a writer myself, I find this rather depressing. I use parallel sentence structure, group things in threes, use unusual-but-accurate words, and come up with my own metaphors because those are good ways to make your point. I'm also inclined to restate and overclarify things to minimize the chance of being misunderstood. I hate the idea of my writing being mistaken for AI slop. At least I type my em-dashes as --, which LLMs don't do.
This is the issue. It's not that this is "LLM writing style". It is just formal writing. The thing is most people write like third graders, so these stick out with good reason, just not that 'it is AI'.
In general there aren't good ways to tell TBH. Literally giving it the command to 'not write like an AI' would make half of these disappear.
Having the AI edit a text for you would add some of these. Not because it is AI, but because that's proper writing.
I don't think it's necessarily just a checklist of things, but rather the way an LLM's output resembles these techniques that puts it into an uncanny valley of writing. As a writer, you use these techniques deliberately and thoughtfully. LLMs can't do that, so the output just feels off.
where do you think they learned it?
Maybe you're actually a llm? And you're slowly learning to disguise that you are by using --
Same. I used to write articles for an industry mag and that's how I was taught to do it.
Some word processing apps automatically convert two dashes to an em dash. I've got that turned off right now.
"em dash" isn't hyphenated, fyi.
That feels like it should be ironic
You can take em—dash out of my cold dead hands
Fuck — as an designer and typography nerd I just love them em-dashes.
I'm sure I've used some software that's auto corrected hyphens to em-dashes too, but I can't remember what.
Outlook does
It’s “em dash” and there shouldn’t be spaces around it. FYI.
I‘m German … in German we connect words with small dashes and put spaces around the larger ones … it’s the habit :)
I hate them. They're fine on posters.
But if I see them multiple times in a block of text, I question the author.
Please don't give this any credit. Nonsense like this is already being used to filter web form submissions for things like job applications.
Source: I applied for a role at a medium-sized company a couple weeks ago and was auto-rejected because my cover letter appeared to be AI-generated. I clicked "back", removed an em-dash, and the form was accepted.
Yeah, having the same problem, but with social media. I've been using em dashes since the nineties, when I found out a pretty dash is just option-dash on my Mac. Switched almost to full time Linux now, still using em dashes. I tried using only minus signs for a time, but muscle memory is hard to beat.
I thought it would be the semicolon, judging from the thumbnail. Now that would piss me off; I love semicolons, it would be unfair if they become the hallmark of LLMs.
I also appreciate the long dash but on mobile keyboard it's so awkward to find that nobody uses it for comments anymore.
Personally, I use the em-dash a lot, but I just type it as --.
I find it's the opposite! I can't find it on OG keyboards which is when I use the double hyphen — on mobile I just long press the dash button to find it
TIL I'm an LLM.
It's not just true, it's a fact — we are all of us llms.
Thought I might be one too, until the "tangents and nonlinear storytelling" as evidence of being human and the scene from Megamind where he goes "being bad is the one thing I'm good at" came to mind.
Welp, TIL that I’m a LLM. I frequently use at like 75% of those in my creative writing. I learned to use all of these by reading books.
That's the issue - so did AI. Real people use them less often because real people often don't read a lot of books. AI might not really understand the words it's spewing out, but it's at least good at formatting them as it's been shown.
The summary isnt hitting the nuance.
I see a lot of AI garbage.
You get emojis as list elements. You see a lot of dry writing, like you're reading a blog. You get vague nothings that sound good, but are meaningless.
Well, so did LLMs haha
For me the most obvious tell is using 16 paragraphs to say something that could have been said with 16 words.
I have ADHD and like en and em dashes. I've been known to use emoji points to make my 16 paragraphs easier to read.
Fortunately I think the constant personal tangents arr saving me
I get what you're saying, but I'll just clarify that my 16 paragraphs vs 16 words was about wordiness, not layout.
That's because you're probably not used to people from STEM areas who tend to be thorough rather than risk that some things might be mis- or not at all understood: the less one is sure about the level of knowledge or ability to keep up of those on the other side, the more thorough the explanation becomes.
Also the deeper you think about something the more elements there are to explain.
I have a STEM background myself and spent a good bit of my career writing (relatively poorly in my opinion) technical documentation. I understand what you're saying and I guess I didn't make my point very well.
I was hoping people would understand that I was referring to the enshitification of internet search results - where every search leads to pages of results of entire articles about very simple topics that say basically nothing. It seems obvious to be, though I admit I'm making an assumption, that the vast majority of these articles are LLM generated fluff attempting to lure people to pages to generate ad revenue.
You just described the last two decades of cooking/recipe websites.
Dr. Ian Malcom: “SEO ruins succinct writing, AI trains on bloated SEO ‘optimized’ text, AI produces bloated SEO-optimized output…”
I find that a more human trait with writing.
People go on and on about nothing and have broken sentence fragments. AI tends to be "too clean".
literally any tell will be fed back into the models to make them harder to detect, I wouldn't be surprised if this is already being automated
Apparently GPT-4.5 has a 73% Turing test pass rate
75% is already huge, 85-90% would just be a marginal change from that. Considering ive been accused of being an AI many times, it leads me to believe that LLMs are already where they need to be in terms of writing format. We just need to work on hallucinations and then AI will start to become actually useful
Oh absolutely. I take a tech-optimist appraoch when discussing new technologies, because the alternative is that everyone outside of the 1% dies in less than 20 years. That seems boring to discuss to me, especially since while i may be a techno-optimist, im also a political-realist. Hence I can see that no one (escept luigi) is actually going to do anything impactful before we all die, so there just doesnt seem to be much point in discussing it. But maybe im wrong, maybe someone will finally decide that maybe we shouldnt be allowing evil people to run the world. In that case the techno-optimist route seems more likely!
Edit: premptive little thought experiment for anyone wanting to disagree. Take whatever counter point you are about to make, and then ask what the results of that "good action" actually were. Likely a lot of nothing, regaurdless of how well intentioned it may have been. Trump is still president after all, and were still just as fucked. The same could be argued about luigi, it doesnt actually seem he achieved much. But hey, he did in fact get rid of one problem directly at the source, which is more than can be said for anyone else in the modern era afaik.
I'm not going to stop using em dashes. Find a different indicator!
Same.
Semicolons are also an indicator. Who the hell uses those?
I use it occasionally, when writing lists with lots of detail, or to separate parts of the sentence where I already used multiple commas.
It's one piece. You tell me I want to spot cars one thing they have is wheels, I'm not going to immediately assume every bike I see is a car. But taken with other signs it builds a pattern.
Honestly I've found the best way to spot LLM is just use it an absolute crap ton. You'll start to be able to spot it the way you can recognise the style of an author or director.
Repeating the same content with different wording later in the text. For instance: saying the same statement you did earlier a little differently
It's like they read a Vonnegut book and said, "That right there is peak human!" And I'd agree!
Good luck, the next rounds of training will iron these out and the line will continue blurring. These tips will flag a lot of false positives from educated people, and those comments are valuable. Maybe if you look for human-style mistakes you will have hope determining if comment was made by a human, like capitalisation, autocomplete issues, and typos.
It's much easier spotting AI in photos, videos and probably audio.
TIL that following many decades of carefully observed writing best practices makes you an AI.
This stuff can also pip up if AI was merely used to spellcheck.
This list would be wrong way too often dude
Not really. Spell check would only correct typos and grammar. It's still your style, your thoughts, your expression with the language.
If you're putting your whole essay into a AI tool, yeah... It's going to turn it into garbage with the list above. And that's on you.
AI is stupid and often does more than told, even if the text is short.
And replacing the hyphen would not be replaced
Well, guess I'm half AI or, to put it another way, I'm similar to a machine that's being frequently affected by cosmic rays when doing calculations
meanwhile here i am with the silly three-em dash⸻too long to be used in most cases, so it's rarely used when training ai⸻and the two-em dash⸺not quite as absurd, but still pretty rare⸺and i keep on looking for weird characters to use regularly
Although I wonder how long these signs would be that effective for, since they seem really model-specific.
Weird that AI can't handle talking in the first person. Why just the other day I, a human, was saying to my sister, who is also human, about how strange AI is. See, she was grew up in another home with her mother where they didnt use as much technology. We shared a father who fought in Desert Storm. His favorite color was blue, like the kind you see in very thick ice. See most people think ice is just clear, but that is only becuase you need a lot of ice to properly refract light into its true blue color. Refraction works because colors come from different wavelengths of electromagnetic radiation, and those wavelengths bend at different angles when refracted by passing through transparent materials.... what was I saying? Oh yeah, AI is strange.
Useful alternative not using a AI detector.
So just add this to your prompt to not do these things? None of the items listed can’t already be handled by existing LLMs.
I have done this with work stuff. Doesn't always work.
We're past the point where it's reasonable to assume ANYTHING viewed onscreen isn't AI generated. The unending content created with AI will be the final nail in the coffin. The screens have subdued us. Hypnotized us into submission. Shaped everything we know and believe. I could probably list a thousand companies, products, trademarks, and slogans. What a waste of cognitive function.