41 Comments
User's avatar
TheOtherKC's avatar

It's not one-to-one, but some of these tendencies were written about almost a century ago by George Orwell in Politics and the English Language, when describing things he saw in poor writing by then-contemporary writers. (https://www.orwellfoundation.com/the-orwell-foundation/orwell/essays-and-other-works/politics-and-the-english-language/)

I don't know if there's enough overlap for me to concoct some grand unified theory or for me to say "there is nothing new under the sun". If there is something deeper, I'm not about to tease it out on this bus ride to the office, but it does seem worth pointing out.

Alberto Romero's avatar

Indeed. There are similarities. But I will say, with all due respect to Orwell, whose writing I admire, that specific essay is quite bad (his analysis is great, his advice is not). I think, unfortunately, that AI writing is even worse than the kind of journalistic and political writing Orwell wrote about.

I.M.J. McInnis's avatar

Agh, accurate and helpful and will share... but damn if I wasn't unsettled and surprised by the conclusion.

Alberto Romero's avatar

Thinking about it I wouldn't say 50% was AI tbh, more like 30-35%. Because the ideas are mine and the tone/style are clearly mine. As are the jokes the quotes... The essence is, I would say, 99% mine. It creates an interesting conundrum nevertheless

Privacat's avatar

I had the same conundrum and it's important to acknowledge. In a recent post, I admitted that Claude meaningfully contributed ideas and, importantly, reminded me of some subtle connections I had started to make but hadn't fully fleshed out. 95% of the post was me, but I benefited from the word-stringing/pattern-recognition as well. Does that make the ideas in the piece worthless? Or worth less?

I don't think so, but for some readers (Ted Gioia, perhaps) it might.

Alberto Romero's avatar

Yeah, I also don't think so. It's perfectly valid. I think that's within the reasonable aspects of AI-human integration

Harold Kildow's avatar

I’m so old I can’t figure out why anyone tries to hide behind an AI for writing. Are you a writer, or an effing “prompt engineer”? A pox on their houses. Also, may it please God to erase prompt engineer from our memories, like “wizard” in the early days of the interwebs.

Ryan Voorhees's avatar

This is the first “spot AI writing” piece I have seen that does not stop at word-level tells and actually talks about structure and intent. The best sections are the treadmill effect and the seesaw sentence pattern, because they describe what it feels like to read the thing, not just which words to watch for. Also appreciated the admission that no single sign is reliable. That is the only honest conclusion.

Alberto Romero's avatar

Thank you Ryan! I agree it's important to go deeper and no sign is perfect

Privacat's avatar

This was great, and I liked the equivocation see-saw point quite a bit. It overlaps with my re-reading of Robert Greene's 48 Laws of Power, and Law 25 in particular.

Personally, this is one of the biggest 'tells' for me of when I'm interacting with an LLM. Or a boring human.

Bring on the spiky and spicy takes!

Thomas Gerbaud, Ph.D's avatar

nice one

actually reading AI text is definitely boring, too well calibrated and boring. and terribly not funny

did I say boring?

James's avatar

Very accurate essay on the matter. I did ask GPT5.2 about its writing style and it admitted that its post training was meant to be nonoffensive and tries to be complete and correct. This is where certain words -- hum, echo, silence -- as well as the negative parallelisms occur. Such being the case, there may be a way at the inference layer to "zero" out the default LLM psychology and introduce human psychology in its place. Promising, but not perfect so far, but much cheaper than fine tuning.

Ricardo Acuña's avatar

👏Very interesting and useful post. In some sense you played a kind of Turing test, where it is very difficult for us to discern if the writer in some parts is a machine or a human. BTW, with all that is going on in AI music, we also need an article about"10 Signs of AI Music That 99% of People Miss". It is amazing how AI authored music already top Billboard and Spotify charts in some genres. So, I´m afraid, could "AI writing" top Substack charts someday without us even realizing it?

Alberto Romero's avatar

Great question and I'm afraid it may be happening already to some degree. I don't think no one can reach that high due to AI but surely with its help!

Ricardo Acuña's avatar

A NYT article: "Why Does A.I. Write Like … That?", scared me with this phrase: "A.I. language is increasingly coming out of human mouths". From a psychological perspective, we will probably end at some extent talking and writing as AI does, the same as we absorb the culture, language accent, art, and so on, from all the people and places around we live on. I find your article to help me as a firewall to detect AI writing and not get caught, at least for some time.

Alberto Romero's avatar

Haha funny that you say that because I've already written another essay starting from that exact paragraph in that NYT's piece (I really like Sam Kriss's writing)

Neill Killgore's avatar

This was great! At some point in the post, I started to wonder if this was partially AI written. I then thought, "nope, not this guy!" Then continued to feel generally uneasy with some of the writing.

The problem is that I don't know why or what is making me feel that way. I can't point it out. I can't pick out specific, concrete examples. I feel like that makes me even more uncomfortable, which is why I continue to read your stuff -- to feel uncomfortable. Because, like vegetables, I'm pretty sure it's good for me.

Alberto Romero's avatar

Hahaha being compared to vegetables is a new one!! Thank you Neill, this was *exactly* my intended reaction. Hope the ideas are useful anyway (they're mine!)

Ido Hartogsohn's avatar

Excellent post, which goes deeply into the essence of AI writing, and you've obviously thought about this long and earnestly. I just wish you'd stop tormenting us with these this-is-actually-AI-writing traps.

And I am curious, since you've often discussed your uneasy relationship with AI writing. To what degree do you use AI in your writing and what is the right place for it in a writer's life, to your mind (if at all)?

Alberto Romero's avatar

Haha I can't! Because this needs to be viscerally understood! (Although I bet you can more or less tell which parts are obviously not AI, I think it lacks quite a few important writing skills, regardless of prompting). I won't use AI except for these kinds of experimental pieces. I like to be familiar with the state of AI writing but it's just not that good. Anything you get from ChatGPT or Gemini is, in my view, utterly unpublishable. Although I see all the time so I guess I belong to the minority. However, I do take editing advice from LLMs as well as suggestions for vocabulary (as English is my second lang). I have less and less grudges toward AI writing as time passes. If people can't tell, then what's the harm actually? I'm not so sure anymore!

Jasmine R's avatar

The harm is that everything is now suspect.

People are growing wary, as people don't want to be fooled or lied to and they want to know what's real and what's not. The more AI-generated content proliferates, the more likely people will hesitate to engage with new writing, etc. It's like how people search for results before 2022 on Google. It's so hard to sift through, so you just avoid it. How can people create and enjoy creative works freely if everyone is worried about whether or not it's real?

Ido Hartogsohn's avatar

That's a good philosophical question. To me it seems AI is good in writing certain types of texts, or in assisting the writing of certain types of texts (it's great at grant proposal writing, for instance), but does not produce particularly interesting texts by itself. There is a cyborg space that emerges between author, AI and text where synergistic collaborations is possible.

AI can erode the sense of meaning for writers, but it can also enhance writing practices. As long as the texts emerging are not worse, but better, I also find it difficult to object. It's still too early to tell the real impact, though.

kipling_sapling's avatar

This is an excellent article and I love that avoiding the tells in my own writing would make me a better writer overall.

J.A. Evans's avatar

Everyone seems to think they can "tell" but when tested, even the best are only slightly better than chance.

LEDPolicy's avatar

Good taxonomy. One addendum: these aren’t just “AI tells,” they’re incentive artifacts. Safety tuning, authority simulation, and verbosity are design choices, not accidents. As long as models are optimized for compliance and completeness, writing will skew abstract, padded, and bloodless—by construction.

Oli G.'s avatar

Aren't AI short emotionally crafted stories already every where (because they drive clicks->traffic->ad revenues) ?

https://open.substack.com/pub/chimeras/p/the-end-of-human-writers-how-synthetical

Ereiklc's avatar

You didnt get the memo about the emdashes :)

Ereiklc's avatar

AI writing tends to use a lot of emdashes because in the training data emdashes signaled “high quality writing”. So by that heuristic I could tell early on that some portion of your piece was AI written, but I don’t hold that against it since the substance is good! This piece describes the emdashes thing more https://www.theintrinsicperspective.com/p/our-overfitted-century

Alberto Romero's avatar

None of the em dashes is from the AI, though. Em dashes is really not a good heuristic. Only *punchline* em dashes

Manoel Galdino's avatar

About latin bias. As a Brazilian, my English is full of Latin words because it's easier to me. To give an example, "paradise" seems more pedestrian than "haven" to me. Though I suppose it's the opposite for native speakers. So, keep that in mind when assessing if some piece is written by a Human or not.

Alberto Romero's avatar

I'm Spanish and I still realize it when it happens to AI. None of these signs should be taken in isolation. AI engages in all of them