Signs of AI Writing

Are we all going to start sounding like this?
Mark Into
April 23, 2026
Sources
Miscellanous

Ten tells of AI writing.

The first quiz tags each highlight with one of four color families: vocabulary, rhetoric, puffery, and stock phrases. Those colors are families, not the signs themselves, so several signs can share puffery or rhetoric.

Puffery
01

Undue emphasis on significance, legacy, and broader trends

LLMs puff up importance by tying arbitrary facts to broader cultural, historical, or symbolic currents. They "shout louder and louder that a portrait shows a uniquely important person while the portrait itself fades from a sharp photograph into a blurry, generic sketch."

Explore more
Puffery
02

Undue emphasis on notability, attribution, and media coverage

LLMs act as if the best way to prove a subject is notable is to hit readers over the head with claims of notability, often by listing sources a subject has been covered in. They may inaccurately attribute superficial analyses to the source.

Explore more
Rhetoric
03

Superficial analyses

AI chatbots insert surface-level interpretation about significance, recognition, or impact, often by tacking an "-ing" phrase onto the end of a sentence.

Explore more
Puffery
04

Promotional and advertisement-like language

Even when prompted to use an encyclopedic tone, LLMs tend toward the prose of a travel brochure or press release. They insert promotional language while claiming to have removed it.

Explore more
Rhetoric
05

Vague attributions and overgeneralization of opinions

AI chatbots attribute opinions or claims to some vague authority, a practice called weasel wording. They also exaggerate the quantity of sources holding a view, presenting one or two citations as widely held positions.

Explore more
Vocabulary
06

"Words to watch" (AI vocabulary)

A distinctive cluster of words appears far more often in post-2022 text than in comparable text from before. One or two occurrences is coincidence; an edit introducing many of them, many times, is one of the strongest tells for AI use.

Explore more
Rhetoric
07

Negative parallelism ("Not X, but Y")

LLMs reach compulsively for constructions that define something by what it isn't. Variants include "It's not ..., it's ...", "no ..., no ..., just ...", and "This dispersal is not dissolution."

Explore more
Rhetoric
08

Rule of three

LLMs overuse the rule of three, from "adjective, adjective, adjective" to "short phrase, short phrase, and short phrase." They use this structure to make superficial analyses appear comprehensive.

Explore more
Opener
09

Outline-like conclusions about challenges and future prospects

Many LLM-generated articles end with a "Challenges" section following a rigid formula: "Despite its [positive words], [subject] faces challenges..." closing with a vaguely positive forecast. Often paired with a "Future Prospects" section.

Explore more
Vocabulary
10

Elegant variation

Generative AI often avoids repeating a word by rotating through synonyms or related labels. The result can sound polished, but it muddies reference: the same person, work, or idea keeps getting renamed instead of simply repeated.

Explore more
Caveats

Age of text relative to ChatGPT launch. ChatGPT was launched to the public on November 30, 2022. Although OpenAI had similarly powerful LLMs before then, they were paid services and not easily accessible or known to lay people. Thus, if an edit was made before November 30, 2022, AI use can be safely ruled out for the corresponding text.

AI detection tools. Do not solely rely on artificial intelligence content detection tools such as GPTZero. While they perform better than random chance, these tools have non-trivial error rates. Detectors can be susceptible to factors such as text modifications, including paraphrasing, markup, and spacing changes, and the use of models not seen during detector training.

Your detection ability. Test your AI detection skills at Wikipedia:AI or not quiz. Do not rely too much on your own judgment. While research on humans' abilities to detect AI-generated text is limited, a 2025 preprint shows that heavy users of LLMs can correctly determine whether an article was generated by AI about 90% of the time, which means that if you are an expert user of LLMs and you tag 10 pages as being AI-generated, you've probably made one false positive. People who don't use LLMs much do only slightly better than random chance, in both directions.

It is also worth noting that writers may adjust their behavior to avoid accusations of AI, or may be defensive about using AI tropes.

A different dialect

You can tell. You've opened a Reddit post, a video, and something about it felt wrong before you could say why. But why?

Why does AI write like that?

It's close to human language, but it's a little off. There's the weird use of the em dash, the overly excitable tone, the weird sentence constructions like "it's not just X, it's Y," or the overuse of certain words like delved, commendable, and meticulous.

AI is speaking a different dialect of language, putting it into a kind of linguistic uncanny valley. It sounds similar to how you talk, but with little incongruities that make the overall thing feel unusual.

That's because of how AI is trained. The chatbot learns from human feedback, and in turn it's always chasing good feedback. That's what makes it sycophantic: humans give good feedback when you suck up to them.

That means the AI has picked up a bunch of linguistic behaviors that make it sound more obsequious. It'll engage in more emotional validation and collaborative communication, which is why it'll say things like "you're absolutely right." It'll use less emotionally charged language, since it's trained to sound more neutral than we irrational human beings. But it'll still lean positive: more approving adjectives, politer speech, because it's engineered to sound like a servant. And it'll speak indirectly, telling you something might be the right or wrong answer, in a way that's less likely to offend.

AI chatbots employ less rhetorical variance than humans and tend to be more formal, so you actually get more overly grammatically correct language. They'll also use every word at a slightly different frequency than humans do, which makes the overall effect sound way more robotic. They stick to these grammatical structures because the structures have proven effective; they got enough positive feedback during the reinforcement learning process to keep showing up.

It's not wrong, it's right.
Extra

Why do AI videos sound like that?

It is something that appears only when you give a prompt to create, but do not give a script. It is something we heard in this class with Professor Nathenson's podcast.

Example discussed in class: Jeremy's AI-video speech pattern clip. Jeremy Finds AI on TikTok

Influencer Speak

Influencer speak is a speech pattern commonly associated with English-speaking digital content creators, particularly on platforms such as TikTok. It is used to maintain audience engagement. This style is characterized by linguistic features such as uptalk, where intonation rises at the end of declarative sentences, and vocal fry, a low, creaky vibration in speech.

hook-heavy uptalk vocal fry engagement cadence

"You might not realize this, but this tiny detail changes everything?"

Both voices are positive-feedback machines. They keep the gestures that seem to work: reassurance, momentum, polish, and a little too much certainty.

Are we all going to start sounding like this?

Probably some of us already do. AI tools are everywhere. Gmail now finishes your sentences. Your boss drafts a memo with it. Students paste essay prompts through AI.

This probably sounds like AI to you. It is not, sadly. Humans copy the language they see.

The "delve" spike is real
PubMed raw counts from Scientific American, with 2024 shown as only the opening months. Kobak et al. separately found "delves" at roughly 25x the pre-LLM trend line.
PubMed papers using "delve" ChatGPT launch — Nov 2022

Kobak et al., writing in Science Advances, went through 15 million PubMed abstracts from 2010 to 2024. "Delves" appeared in roughly 25 times as many 2024 papers as the pre-LLM trend line predicted. "Showcasing" and "underscores" were close to ninefold.

Other databases show the same bend in the line. Kousha and Thelwall tracked twelve LLM-associated words across six scholarly databases: between 2022 and 2024, delve rose about 1,500%, underscore about 1,000%, and intricate about 700%. Full-text data is stranger still: in PMC, the share of papers using "underscore" six or more times grew by more than 10,000% from 2022 to 2025.

Most new web articles may now be AI-written
Graphite analyzed 65,000 English-language articles from January 2020 through May 2025. Production rose toward 50-50, but AI content appears less often in search and citations.
AI-generated articles AI share of what is surfaced Human share
The exposure loop iii.

Graphite found that before ChatGPT, roughly 10% of new online content was AI-generated. By late 2024, the share had climbed over 40% and briefly surpassed human-written articles before settling near 50-50. The most recent figure cited in the coverage is 52% AI-generated, 48% human.

Most of us are reading AI-assisted articles. But volume is not the same thing as attention. Only 14% of content ranking in Google searches was AI-generated, and ChatGPT cited human-written articles 82% of the time. AI slop is being produced in volume, but not read in proportion.

At the very least, since we see it a lot, we can mimic it.

So yes, some of us will drift toward the model. We see the language, borrow the rhythm, and slowly mistake the average for the normal. But the tells also create a reaction: readers get faster at recognizing machine prose, and writers who care about voice learn to put the friction back in.

Tweaks Adjust