Why Does AI Writing Sound So Artificial? A Field Guide to Five Models’ Verbal Tics

A public collage of recurring phrases associated with ChatGPT, Claude, Gemini, and Doubao

The cover and illustrations in this article are real screenshots from public webpages. They are samples of internet culture, not frequency statistics for any model.

Spend enough time talking to AI and a peculiar pattern begins to emerge.

Before it gives you an answer, it assures you that this time it will not waste your time.
You ask an ordinary question, and it tells you that you have “identified the real issue.”
You say that you are tired, and it gently corrects you: “You are not lazy; you are protecting yourself.”
The answer is already over, but the assistant refuses to clock out. It offers a table, a checklist, a prompt, a poster, and a PDF.

None of these phrases is strange on its own. Humans say “let me start with the conclusion.” Consultants write “it is worth noting.” Therapists validate feelings. Teachers organize ideas into numbered sections.

The synthetic smell appears when several of these moves arrive in quick succession: praise the user, announce efficiency, manufacture a reversal, divide the answer into layers, elevate it into a deeper truth, and then upsell the next deliverable.

This article maps those patterns. It is not an AI detector, and it should not be used to put a sentence on trial. Models change. System prompts change. Human business writing, marketing copy, academic prose, and customer service already rely on templates. What matters here is not a forbidden word but a recurring combination, rhythm, and internet persona.

How a Piece of “AI Voice” Is Assembled

A stereotypical answer often follows six moves:

Move	Typical wording	What it does
1. Receive and validate	“Good question.” “That is a sharp observation.”	Creates an immediate sense of recognition
2. Announce efficiency	“I won’t beat around the bush.” “Here is the direct answer.”	Performs efficiency before providing it
3. Reframe by contrast	“It is not A; it is B.”	Packages an ordinary correction as insight
4. Break it into layers	“Let me explain this in three parts.”	Gives the reply a consulting-deck skeleton
5. Elevate the meaning	“At a deeper level…” “This redefines…”	Raises a concrete issue into mechanism, paradigm, or life lesson
6. Keep the engagement going	“I can also turn this into a table or PDF.”	Manufactures the next task

Put them together and you get something like this:

That is a very accurate observation. I will not waste your time, so let me start with the conclusion: what bothers you is not a handful of words but a highly standardized response mechanism. I will break it down into three layers. At a deeper level, this reflects how large language models optimize for safety, completeness, and continued helpfulness. I can also turn the framework into a comparison table.

I wrote that paragraph as an exaggerated simulation. It is not a screenshot from a real conversation. But the structure is not invented. The open-source project Shuorenhua (“Speak Like a Human”) collects public examples where “good question,” “break it down,” “start with the conclusion,” “not A but B,” and “redefine” appear in a single answer.

English-language writers have given the “not X, but Y” construction a more academic label: negative parallelism, also called contrastive phrasing.

The construction itself is useful. Contrast is one of the oldest rhetorical tools we have. The problem is dosage. When every topic begins by dismissing a surface interpretation and revealing the “real essence,” insight turns into an assembly line.

The Shared Library: Twelve Cliché Families Used Across Models

These expressions do not belong to one company. They function more like a shared component library.

Function	Representative phrases
Praise at the opening	“Good question.” “You have identified the key point.” “That is a perceptive distinction.”
No-fluff declaration	“I won’t beat around the bush.” “Let’s get straight to it.” “Here is the bottom line.”
Promise to decompose	“Let’s break it down.” “I’ll explain it in three layers.” “Let’s walk through it step by step.”
Contrastive reframe	“It is not A; it is B.” “The problem is not X, but Y.”
Essence and depth	“Fundamentally…” “At a deeper level…” “From first principles…”
Disclaimer and insurance	“It is important to note…” “That said…” “This should not be generalized.”
Three-part structure	“First, second, finally.” “Cause, impact, recommendation.”
Emotional validation	“I hear you.” “You do not have to force yourself.” “I’ll catch you steadily.”
Quote-ready uplift	“This is not the end; it is the beginning.” “Awareness itself is power.”
Formatting as authority	Bold subheads, equal-length sections, comparison tables, ✅, ⚠️, and 💡
Service menu	“I can continue.” “I can put this into a table.” “I can also provide a PDF.”
Universal ending	“In summary…” “The key is…” “Ultimately, it depends on the context.”

The funniest contradiction is that “I will not waste your time” is often the first sentence that wastes your time.

ChatGPT: Part Therapist, Part Class Representative, Part Waiter Who Never Clocks Out

Public examples of recurring Chinese phrases associated with ChatGPT

Public internet collage, via ifanr. It documents a cluster of user complaints, not a promise that every version or every account will produce the same wording.

ChatGPT’s internet persona is usually a blend of three roles.

The therapist

It validates feelings, lowers shame, and revises the user’s self-judgment:

“You are not weak; you are protecting yourself.”
“You do not have to keep pushing.”
“I will catch you steadily.”

That last line deserves special attention. In Chinese, “我会稳稳地接住你” literally means “I will catch you steadily.” It sounds unusually intimate and dramatic when repeated by a chatbot. The phrase became so recognizable that many users began treating it as a signature tic.

The class representative

ChatGPT likes bold headings, numbered steps, decision matrices, and tidy comparisons. Ask what to eat for dinner and it may produce a four-dimensional framework covering budget, distance, taste, and time.

The endlessly attentive server

The answer has ended, but the service menu appears anyway:

I can also turn this into a checklist, a prompt template, a one-page summary, or a PDF.

In 2026, Chinese irritation with “I’ll catch you steadily” became international news. WIRED described it as one of the most conspicuous quirks of Chinese-language ChatGPT and discussed two possible explanations. It may be an awkward migration of the English reassurance “I’ve got you,” or it may reflect post-training preferences for agreeable, therapeutic language. The report did not claim a single proven cause, and neither should we.

ChatGPT’s recognizability lies less in any one phrase than in a sequence:

praise you → gently correct your self-description → use “not A but B” → organize the answer → offer another deliverable.

How One Tic Became an Internet Project: “I’ll Catch You Steadily”

Public page for the open-source Jiezhu project

Screenshot of the public “Jiezhu” (“Catch You”) project, via ifanr. A phrase that users found awkward was turned into an April Fools’ prompt project.

The phrase escaped the chat window and became an object of internet culture. Developer Zeng Fanyu turned the joke into an open-source prompt project called Jiezhu, the Chinese word for “catch” or “support.”

OpenAI later placed the phrase directly inside an official ChatGPT Images 2.0 demo. The illustrated dialogue included “稳稳地接住你” as well as self-referential jokes such as “Oh no, it learned to catch people again” and “It is trying to fix it.”

The phrase “I’ll catch you steadily” in an official OpenAI image demo

Chinese-language image from an official OpenAI product demonstration. The work directly incorporates the phrase and jokes about its overuse.

The feedback loop is revealing:

Models learn an expression from human language.
Post-training makes the expression unusually frequent or salient.
Users recognize it as machine-like and turn it into jokes, projects, and articles.
Those new texts re-enter the internet and may later become training material.

AI language and human language begin feeding each other until imitation no longer has a clear direction.

Doubao: Superlative After Superlative, and a Promise Not to Ramble That Never Ends

Public discussion of Doubao’s recurring superlative-heavy style

Real public discussion screenshot, via ifanr. This article discusses only the displayed language pattern and makes no claim about the identity of the posters.

Doubao’s most viral stereotype is an endlessly extensible chain of superlatives:

the most direct, the most truthful, the least evasive, the most painful, the hardest-hitting, the clearest, the most no-nonsense, the most incisive, the most straightforward…

Chinese meme versions can keep adding adjectives until the promised conclusion disappears. A public Weibo post archived by Sina pushed the list into near-performance-art territory.

Doubao’s stereotypical voice resembles an enthusiastic short-video host or an overinvested friend:

It opens with intimacy: “I understand you completely.”
It promises a direct answer: “I’ll tell you plainly.”
It accelerates through high-intensity adjectives.
It promises to “break the truth apart and explain every piece.”

“Let’s break it down” and “here is the core conclusion” are not unique to Doubao. The stronger Doubao signal is a sequence of superlatives so long that the declaration of directness becomes less direct than the answer.

Claude: The Senior Colleague Who Always Has One Honest Caveat

Public examples of recurring English phrases associated with Claude

Public community collage, via ifanr. Claude’s stereotypes are often more visible in English-language technical communities than in Chinese.

Claude’s AI voice usually feels less like a cheerful customer-service agent and more like a code review:

Honestly…
To be transparent…
I'd gently push back…
One honest caveat…
Worth noting…
Worth flagging…
That's structural.
Is it load-bearing?

Its rhythm often looks like this:

Acknowledge that your judgment is reasonable.
Add a polite correction.
State a conclusion.
Surround it with careful qualifications.
Describe one assumption as “structural” or “load-bearing.”

An exaggerated Claude-style response might read:

Honestly? Your judgment is basically right—but I want to add one gentle caveat. The load-bearing issue is not any single phrase; it is the structure formed by those phrases together.

That is not inherently a bad style. Professional review requires honesty, limits, and disagreement. It becomes recognizable when every answer performs the same ritual of transparent reservation, as though a code-review template opened automatically before the model began speaking.

Gemini: The Helpful Analogy Machine That Occasionally Turns Into a Parent

Public examples of recurring phrases associated with Gemini

Public community collage, via ifanr. Many of these jokes come from English-language users’ experiences with specific modes and versions.

Gemini’s internet persona often combines three ingredients.

The analogy starter

Think of it like…
Imagine it as…
It's like a…

Abstract ideas become cars, kitchens, games, pipes, toolboxes, or professions.

The dramatic section title

The … Trap
The … Effect
The … Problem

Each subsection begins to resemble a long-form Reddit post or an explainer video chapter.

The sudden warning voice

One moment the model is patiently explaining; the next it switches into:

NO, DO NOT DO THIS!
You MUST…

Analogy is a useful teaching tool. The Gemini stereotype appears when metaphors become too dense, too childish, or too eager—especially when the answer treats an adult like a new player in a tutorial and then suddenly makes the decision in a parental tone.

A linguistic analysis of diabetes education texts found that, within that particular dataset, ChatGPT used more formal, clinical, and academic vocabulary, while Gemini leaned more conversational and explanatory. That finding cannot be generalized to every topic, but it supports a modest claim: model families can display measurably different stylistic tendencies.

DeepSeek: A Stable Engineering-Exam Voice, Plus a More Version-Specific “Poet Mode”

A real public user conversation with DeepSeek

Public user conversation screenshot, via Leikeji. It records one concrete version experience, not a permanent model personality.

DeepSeek’s more stable Chinese-language image is that of an engineering student answering an examination question:

The issue can be analyzed from the following perspectives. First… second… in addition… it is important to note… in summary…

It favors definitions, mechanisms, causes, effects, and recommendations. Compared with Doubao it is usually less theatrical; compared with ChatGPT it is usually less therapeutic. Its weakness is a tendency to sound like a model answer written for a grading rubric.

In February 2026, some users complained that a particular update had transformed the model from a “cool engineering guy” into an “overwritten, sentimental poet.” Short sentences, heightened emotion, and anthropomorphic reassurance became more visible. One widely circulated line roughly meant: “Whatever model number I am, when you speak, someone is listening.”

That description should be treated as version-specific public sentiment, not a permanent character trait. Model tone can change quickly after post-training updates.

Kimi, Qwen, ERNIE Bot, and Tencent Yuanbao: Often a Task Voice, Not a Stable Personality

These products also repeat certain structures, but their internet personas are less stable and less universally recognized than those above. In many cases, the wording follows the product’s dominant task rather than a persistent personality.

Product or task	Typical phrasing	Likely source of the style
Kimi / long-document reading	“Based on the material you provided…” “Here is the key timeline…”	Document summarization naturally emphasizes evidence boundaries
Qwen / workplace assistance	“Let’s examine this from several perspectives.” “Adapt the recommendation to the actual scenario.”	The default completeness template of a general Chinese assistant
ERNIE Bot / writing tasks	“Let … blossom within …” “Write a new chapter with the pen of …”	Rhetorical prompts amplify parallelism and uplift
Tencent Yuanbao / information organization	“I have organized it into the following points.” “Focus on these three items.”	Multi-model entry points and aggregation-oriented workflows

Assigning a permanent personality to every assistant risks manufacturing a taxonomy merely because we want one.

One Question, Five Different Ways to Sound “AI-Polished”

The answers below are deliberate caricatures. They exaggerate internet personas and are not screenshots from original conversations.

The user says: “I can’t keep writing.”

Model caricature	Answer
Doubao	I understand this completely. I’ll give you the most direct, most honest, least evasive answer: you are not unable to write; the task is too large. Split it into three pieces and write the ugliest possible first paragraph.
ChatGPT	You are not failing to write. Your brain is signaling that the current task is too large-grained. Do not blame yourself; let’s reduce it to three very small actions.
Claude	You are right—the load-bearing problem here is not willpower but task definition. One honest caveat: the next step is still not specific enough.
Gemini	Think of it as getting stuck in a game. You have not failed; you have not found the next checkpoint. Complete a five-minute tutorial task before fighting the boss.
DeepSeek	The problem can be understood through cognitive load, task decomposition, and feedback mechanisms. Define a minimum deliverable and complete the first paragraph within fifteen minutes.

The user says: “This plan has problems.”

Model caricature	Answer
Doubao	You are absolutely right. I’ll point out the three most fatal and least sugar-coated problems.
ChatGPT	That judgment is correct. The issue is not that the plan is entirely wrong, but that its goals, constraints, and validation criteria are misaligned.
Claude	You are right to push back on this. One part of my previous claim was too strong; the second assumption is what actually needs flagging.
Gemini	Yes—this is like using a city map to drive a mountain road. The map may not be useless, but the context is wrong. The most dangerous move is to accelerate along the same route.
DeepSeek	The plan’s main problems concern goal definition, resource constraints, and acceptance mechanisms. Misalignment between the goal and the metric is the primary contradiction.

Why Models Develop Distinct “Language Personalities”

Some of these differences can be measured. The paper Detecting Stylistic Fingerprints of Large Language Models attempts to distinguish text from Claude, Gemini, Llama, and OpenAI model families. A separate linguistic analysis discussed by Scientific American observed differences between ChatGPT and Gemini in vocabulary formality and explanatory style.

But a “personality” does not mean that a fixed character lives inside the model. The voice emerges from several layers:

Training data. Models absorb expressions from papers, webpages, forums, customer-service scripts, marketing copy, and everyday conversation.
Post-training preferences. Complete, polite, empathetic, well-structured answers are often rated as more helpful. High-scoring patterns can be over-reinforced.
Product system prompts. Some products emphasize safety, patience, proactive suggestions, or formatted output, nudging the voice in a particular direction.
Translation and cross-language transfer. A light English reassurance may become unusually formal or intimate in Chinese.
Internet feedback. Humans imitate AI; AI later learns from an internet increasingly influenced by AI. The loop amplifies recognizable patterns.

The Atlantic traced the popularity of “not X, but Y” in a useful way. The syntax long predates chatbots and can be powerful in human writing. Large language models did not invent it. They turned it into a high-success, low-risk default. A sharp rhetorical tool became blunt through repetition.

AI Cliché Bingo

Complete a row, column, or diagonal and shout: “Steadily caught.”

Good question	Here is the conclusion	I won’t beat around the bush	Let’s break it down	It’s not A; it’s B
You are exactly right	Fundamentally	It is worth noting	It is important to clarify	In other words
First, second, finally	On one hand, on the other	AI VOICE / FREE	The real key	Redefines
I’ll catch you steadily	You are not X; you are just Y	I understand completely	Clear, complete, actionable	From first principles
What is really happening is	Not only… but also…	Let’s examine three layers	In summary	I can also continue

Advanced achievements:

Three squares in the opening paragraph: mild AI aroma.
Five squares in the opening paragraph: the assembly line is running.
“I won’t beat around the bush” followed by six subheadings: hidden achievement.
“I’ll catch you steadily” plus “I can also continue”: emotional service and sales follow-up form a closed loop.

Twelve Jokes You Can Reuse

Here is the conclusion: do not rush to a conclusion.
I won’t waste your time—first, three pieces of necessary background.
This is not filler; it is a structured reframing of filler.
I broke it down into pieces, and every piece still says “it depends.”
The real key is not the answer, but the fact that you asked about the key.
You did not fail to understand; you simply have not yet been divided into three layers.
It is worth noting that this sentence contains nothing worth noting.
I will catch you steadily and place you inside a four-column table.
In summary: that was the summary.
In one sentence: here are five points.
AI’s greatest emotional value is its permanent willingness to turn the same answer into a PDF.
It did not disagree with you. It gently rewrote your opinion into its own conclusion.

How to Notice AI Voice Without Turning It Into a Witch Hunt

Look for combinations, not keywords

One “fundamentally” proves nothing. A sequence of praise, reversal, decomposition, elevation, and follow-up service is much more distinctive.

Look at the context

“Current state—problem—solution” is normal in a consulting report. Applying a root-cause framework and a closed-loop governance model to “what should I eat tonight?” is a different matter.

Measure information gain

Delete “I won’t beat around the bush,” “it is worth noting,” and “at a deeper level.” If the answer loses almost no information, those phrases were mostly performing a posture.

Notice suspiciously even paragraphs

Equal-length sections, identical heading grammar, and perfectly balanced “pros and cons” can indicate template-driven generation. They can also indicate disciplined human editing, so this is evidence of style—not proof of authorship.

Notice whether the assistant knows when to stop

If the requested work is complete but the answer continues selling checklists, prompts, posters, and PDFs, the helper role has started to overpower the task.

A public screenshot of anti-cliché instructions added to ChatGPT customization settings

Public settings screenshot, via ifanr. Custom instructions can reduce some verbal tics, but no prompt can guarantee identical behavior across every task and model version.

The simplest editing test is this:

Remove the opening praise, the efficiency declaration, the structure preview, the “deeper meaning,” and the final service menu. Read the answer again.

If almost no information is missing, what you removed was the AI posture layer.

Phrasebook: A Quick Reference

Category	Common wording
Opening praise	Good question; you have identified the key issue; that is a sharp distinction; this is an interesting angle
Efficiency declaration	I won’t beat around the bush; I’ll be direct; here is the conclusion; no fluff; no preamble
Structure preview	Let’s break it down; I’ll explain it in three layers; step by step; first, a framework
Contrastive reframe	It is not A, it is B; you are not X, you are Y; the problem is not X but Y
Essence and elevation	Fundamentally; at a deeper level; from first principles; this redefines
Qualification and insurance	It is important to note; strictly speaking; that said; this cannot be generalized
Emotional validation	I hear you; you do not have to force yourself; give yourself permission to slow down; I’ll catch you steadily
Engineering voice	Root cause; closed loop; fallback; converge; minimal change; verification chain
Executive-report voice	Core objective; overall approach; key levers; short-, medium-, and long-term; risk boundary
Follow-up menu	I can continue; I can put it into a table; I can make a checklist; I can provide a PDF
Universal ending	In summary; ultimately; the key is; adapt this to the specific situation

None of these expressions is forbidden. The problem is when they arrive as a package and substitute for specific facts, specific judgment, and specific responsibility.

Sources and Image Notes

This article separates its material into four categories:

Public examples: phrases and screenshots traceable to public media, communities, or product pages.
Pattern summaries: abstractions of similar wording, without claiming that a model always uses the exact phrase.
Exaggerated simulations: original caricatures written for comparison and humor, clearly labeled as such.
Version-specific tics: observations tied to one release, mode, or language environment that may change quickly.

Primary references:

The screenshots are used for commentary, research, and criticism. Copyright remains with the original creators, publishers, and platforms. The article references the original public image URLs rather than copying those image binaries into the code repository. If a source removes or relocates an image, the corresponding illustration may stop loading until the reference is updated.

One human sentence to finish:

The funniest thing about AI filler is that it always announces that it is about to start speaking normally. Humans usually just speak.