On July 14, 2026, programmer Johanna Larsson published a blog post you could finish in under two minutes. She wrote a small script that automatically replaced the mind-numbingly repetitive words her AI coding assistant favored — “load-bearing,” “honest take,” “you’re absolutely right” — with absurd, silly substitutes. This featherweight technical post detonated on Hacker News with 405 upvotes and 464 comments, and the comment section veered entirely away from the technology itself — people started telling stories of being “infected in reverse” by AI.
One comment read:
“I stopped using that AI a long time ago, but my colleagues all use it. I read their documents and noticed the word ‘load-bearing,’ thought it was actually kind of useful, and started using it in everyday conversation. Until someone told me: ‘You sound more and more like Claude.’ I never use that word now.”
That comment racked up a huge number of upvotes. Because the person who wrote it wasn’t alone.
How Does a Word Go “Person to Person”?
“Load-bearing” is originally an architectural term meaning “supporting weight” — like a load-bearing wall. When AI uses it to describe “critical logic” or “the part you can’t delete” in code, it’s essentially making an analogy, which isn’t wrong. The problem is frequency.
In the comment section of that HN post, someone kept a record: their AI assistant had settled into a fixed vocabulary in recent conversations — “projection,” “strand,” “frontier,” “quiescence,” “honest,” “residuals,” “rescission,” and “supersession.” None of these words is problematic in itself, but when AI repeats them in every reply, they form a kind of “linguistic fingerprint” — you don’t need to see the author’s name; the word choice alone tells you who wrote it.
This started as just one engineer’s annoyance. What escalated it was the comment section’s second thread: “person-to-person transmission.”
More than one person reported a similar experience: they never used AI directly, but because colleagues use it, partners use it, industry reports use it, these AI-frequent words seeped into their vocabulary through documents, emails, and meeting notes. A commenter who called themselves a “former professional writer” said they wrote a thank-you note to a colleague in collaboration software, and half the people assumed it was AI-generated — “they said I never write anything longer than two sentences, so anything even slightly eloquent couldn’t possibly be human.”
Another commenter got more specific: “I read a book and found AI’s favorite phrases everywhere. I was about to declare it AI-ghostwritten, then I checked the publication year: 2019.” Back then, today’s dominant chatbots hadn’t even launched.
Why Does AI Have Verbal Tics?
The answer to this question is more concrete than you’d think.
Take the word “honest.” A Hacker News user traced it back and found that one AI’s training material included a core document called the “Constitution,” in which “honest” and its variants appeared 57 times. In other words, the AI “learned” to modify its own judgments with “honest” — the root of this behavior is the weight distribution in the training data. That core document contained 57 instances of “honest” and its variants, and the model is probabilistically pushed in that direction: using “honest” is the safest, most human-acceptable choice.
The same logic applies to all AI-frequent words. “Delve,” “tapestry,” “crucial,” “underscore,” “moreover,” “landscape” — according to a 2026 statistical analysis, AI uses these words 50 to 269 times more often than human writers.
This phenomenon can be measured precisely. A language model is fundamentally a probability predictor trained on massive amounts of human text — it selects “the word most probable in a similar context.” When a model generates tens of billions of tokens (semantic units) every day, its tiny internal probabilistic preferences get amplified at the output end into a glaring linguistic homogenization.
One commenter summed it up sharply: “A person has their own linguistic preferences, writes 5,000 words a day, and nobody finds it strange. But an AI model’s preferences get multiplied by ten billion in output daily — any preference becomes a louse on a bald head.”
Key Evidence: Humans Really Are Being “Trained” by AI
In August 2025, a peer-reviewed study from Florida State University first confirmed with empirical data what many had vaguely feared. The research team analyzed changes in word-frequency in everyday human speech before and after ChatGPT’s launch, and the results pointed in a clear direction: AI’s frequent words are seeping into real human conversation.
Specifically, they found that the use of “underscore” rose measurably after ChatGPT’s launch, while its synonym “accentuate” did not. If this were natural language evolution — like “geili” replacing “lihai” — synonyms should rise together or at least show similar trends. But the actual data didn’t look like that. Only the specific AI-preferred word went up.
The researchers named this phenomenon the “seep-in effect.” When Newsweek reported on the study, it quoted a behavioral analyst’s warning: what people should worry about most is “the disappearance of individuality.”
Another study from the Max Planck Institute focused on academic YouTube content creators. They found that in the 18 months after ChatGPT’s launch, these creators’ use of words like “meticulous,” “adept,” and “delve” rose by 51%. The researchers noted that most people aren’t even aware they’re using these words — because individuals can’t see the larger-scale shifts in language patterns.
It’s a bit like boiling a frog. You won’t wake up one morning and suddenly decide to start saying “underscore,” but when every article you read, every video caption you see, and every work email you receive uses the word at high frequency, your vocabulary shifts quietly. The human language-learning mechanism — imitation — is being hijacked by the sheer scale of AI output.
The Debate: Pollution, or a Good Thing?
It’s not entirely one-sided.
These words are often good writing habits in themselves — “delve into” is more precise than “look into,” “underscore” more formal than “say again.” The problem is the sensory fatigue from overuse: like a good song looped 500 times, you just want to smash the speaker.
Other commenters pointed out that many so-called “AI tics” existed before AI, in corporate white papers, management-consulting reports, and academic writing styles. AI merely amplified patterns that were already high-frequency to an uncomfortable degree. Someone recalled that before “load-bearing,” the corporate world had popular metaphors like “stove pipe” and “silo” — all used to death before being replaced.
In other words, AI didn’t invent a new language — it just accelerated the metabolic cycle of linguistic fashion. When a person repeats a phrasing, it’s “personal style”; when an AI repeats it, it’s “data pollution.” The only difference is scale.
But looked at from the other side, scale itself is the heart of the problem. One commenter wrote: “When I see 13 ‘load-bearing’ dashes on the first page of a requirements doc, I know it’s going to be a bad day.” Behind this annoyance lies a layer of signal judgment: when you see these signature words, you instantly realize no one behind the text is actually thinking — it’s just been assembled.
We’re Entering an Era of “Mutual Linguistic Domestication”
What really struck a nerve in this discussion isn’t that AI has verbal tics — that was never news. What made people’s stomachs tighten was realizing they themselves are becoming the trained object.
A Hacker News commenter described an unsettling self-observation: having noticed that AI gives better replies when he swears at it, he developed a habit of cursing at AI. The habit generalized until he had to consciously remind himself not to swear when buying coffee. “Even just writing this experience down,” he wrote, “I can hardly avoid throwing in a few F-bombs to emphasize how absurd this is.”
But it’s not one-directional. There’s a two-way training process between humans and AI. Humans train AI to act more human through feedback mechanisms (upvotes, rewrites, choosing replies); AI trains humans to act more like AI through its ubiquitous output. One comment astutely predicted this: “If every day a popular model repeats ‘load-bearing’ to every developer, eventually developers — especially newcomers who don’t know it’s an AI tic — will start saying it too.”
And what we’re seeing now: that prediction has come true. Developers are the first hit; marketing people writing reports, admins taking meeting notes, and students writing course papers follow close behind. AI’s linguistic patterns are slowly and irreversibly reshaping how we express ourselves, through a path of “documents infecting documents, people infecting people.”
So, What Should We Do?
This doesn’t need to be “solved,” but it needs to be “noticed.”
VICE wrote in one report: “AI is sanding the rough edges off human communication, erasing the tiny linguistic differences that distinguish one person from another, making us sound more and more like the same person — over-polished, unsettlingly enthusiastic, inauthentic human replicas.”
But some see the other side of the coin. The words AI overuses — “honest,” “underscore,” “delve” — placed in any writing guide, are precise expressions recommended for use. They became “tics” for one reason only: they’re used too much. This actually points to a time-worn writing principle: use good words, but use them where they count.
One Hacker News commenter said his current coping strategy is to consciously use the word “I” more in his writing — because AI, until explicitly asked, rarely volunteers first-person voice. This simple trick lets him maintain writing quality while stamping a subtle “human watermark” on his text.
What I want to say is: language was never a fixed, unchanging system. The internet changed how we talk (“hahaha” replaced “laughing myself to death”), input methods changed how we write (pinyin prediction makes certain words easier to select), and AI is merely the latest link in this long chain. What’s different this time is the speed and scale — and an easily overlooked fact: this time, the tool is reshaping how you use it, in reverse.
Realizing this is the first step to change.
Reference Links
- Johanna Larsson: How to stop Claude from saying load-bearing (personal tech blog)
- Hacker News discussion thread
- On-screen and now IRL: FSU researchers find evidence of ChatGPT buzzwords turning up in everyday speech — Florida State University News
- AI Is Changing How We Speak — Newsweek
- AI Is Changing the Way Humans Speak to Each Other — VICE
- Delving into the load-bearing tapestry of AI’s overused words — Jake Orlowitz / Medium
- Wikipedia: Signs of AI writing
- 50 Words AI Overuses (And What to Write Instead) — HumanizeThisAI
- Max Planck Institute: study on language shifts among academic YouTubers after ChatGPT’s launch
Source: jola.dev blog, showing the effect after AI-frequent words were replaced by the script
Source: Florida State University College of Arts and Sciences, Adobe Stock image, FSU study on how ChatGPT affects human speech