Break down the six thousand words sentence by sentence · Why it is not like any version of Sun Yuchen
August 29, 2026
On the night of August 27, Sun Yuchen, who was in Bhutan, released a six thousand word article titled "My Girlfriend Jing Tian." He later told the media that it took him over ten hours to write, occupying the early mornings of two nights, with almost no sleep.
This article does not touch on the case details. The case has been filed, and Jing Tian's side has raised objections to jurisdiction, which has not yet entered substantive hearing. Jing Tian's studio denied the rumors. Who is right and who is wrong will be left to the court.
I will only do one thing: break down these six thousand words, point out which sentences are machine-written, and explain why.
Written in advance
It must be made clear: all AI detectors on the market are unreliable, and I do not have his conversation records. What follows are all textual evidence and inferences about the writing mechanism, not technical identification, nor does it constitute factual accusations.
01
Place the conclusion here first
The material is his, but the sentences are not his.
To be more precise: he provided the memories, emotions, structural intentions, and final decisions, while AI provided the language. This is not "AI ghostwriting," it is "AI co-writing."
One piece of evidence lies in his own statement. Six thousand words, for a winner of a new concept essay competition, would not take over ten hours to write by hand. But the process of "generate—select—splice—polish" is very standard for ten hours.
02
The foundation of judgment: What AI leaves is not wrong, but "too right"
First, let me clarify what I am looking at.
Generating text with a large model essentially involves selecting the word with the highest probability at each step. It naturally follows the smoothest slope. This leads to three consequences, which are the foundation of all judgments in this article.
First, sentences are good, but they are uniformly good. Human writing has a peak information density—some paragraphs suddenly collapse, while others suddenly shine. Machines produce a flat line.
Second, structure is perfect, because it has perfect memory. The entire draft is within its context window; keywords buried in previous text can be retrieved at zero cost five thousand words later. Humans rely on revisions to do this.
Third, instructions are executed one hundred percent. If you tell it "do not write emotions directly," "do not use conjunctions," "use short sentences," it will execute it flawlessly throughout six thousand words. Humans can only execute about seventy percent; the remaining thirty percent tends to be missed—and that missed part is precisely the human touch.
The following ten instances are all specific manifestations of these three points.
03
Break down sentence by sentence
One | The opening "weighing"
The weight of an egg is three point five micrograms. / The weight of fifty million dollars in cash is two point five tons. / … What I want is the latter, mortgaging the former.
Why AI?
Two quantities are placed side by side, and then "the latter / former" is used to close. This is a designed opening device, not the beginning of a reminiscence. People do not make unit conversions when their emotions are at their peak.
A harder point lies in the numbers themselves. Fifty million dollars, if in hundred-dollar bills, is five hundred thousand bills, about one gram each, totaling about five hundred kilograms, half a ton. To reach two point five tons, it would have to be calculated at twenty-dollar bills—two million five hundred thousand bills. (This is my estimation; the original text did not specify the denominations.) A person writing by feel would not arrive at this number; it requires a calculation, and once done, the premise is forgotten.
Technically speaking
The second sentence is almost a forced continuation of the first. Once the model generates the phrase "the weight of X is a certain number," the probability of a similarly structured next sentence skyrockets—parallel structure is self-reinforcing.
And "the latter / former" is the easiest way to close—it only needs to refer back to the previous text, without generating any new information.
Two | A Versailles without a queen
A Versailles without a queen.
Why AI?
After describing the scene of a cleaner sweeping a vacant room, it leaves a blank line, drops this noun phrase without a verb, and immediately moves to the next section. This sentence is well written, but the problem is that it is too good, too independent, and too cleverly placed.
Technically speaking
In the training corpus, literary paragraphs often conclude with metaphors or maxims. The model learns a very strong correlation—that at the position of "scene ends," the next high probability sequence is "abstract summary sentence." So it closes like this every time. Such sentences are sparse and randomly positioned in human writing; in this article, they are as punctual as a metronome.
Additionally, "Versailles" in the Chinese context is a pun for "showing off." This cheap pun fits in perfectly here, it’s just too coincidental.
Three | "Nothing happens"
I suddenly realized that Jing Tian likes a place because it is empty because of her. Nothing happens.
Why AI?
"I suddenly realized" is an epiphany marker, used specifically to switch narrative to thematic statement. It is immediately followed by two sentences starting with "Jing Tian." Then it buries four words: nothing happens.
Six thousand words later, the last sentence of the entire text is: nothing has happened.
Technically speaking
Embedding and retrieval of thematic words is a revision process for humans but a zero-cost operation for the model— the entire draft is always within the context window, recovering a phrase from five thousand words ago is as easy as recovering the previous sentence. The difficulty of writing is completely asymmetric here. The cleaner the appearance of such a long-distance antithesis, the less it resembles a first draft.
Four | The most beautiful sentence, yet the least human
I completely believe what she says. / But I do not believe a word.
Why AI?
The two sentences have similar word counts, and their semantics are strictly symmetrical, with the reversal occurring after the period. Seamlessly fitted.
Technically speaking
The model has a strong preference for generating paired sentences in the format of "semantic antonym + syntactic isomorphism" because these two constraints narrow the distribution of candidate words simultaneously, making it the most comfortable path. Humans write reversals unbalancedly—either the first half drags while the latter suddenly shortens, or vice versa. Here, the lengths are the same and stable.
Five | Two girls with tape
They found that tape applied after three o'clock in the afternoon comes off most easily because that’s when the glass is hottest; later they switched to applying it in the morning. No one taught them this; they figured it out themselves.
Why AI?
The biggest flaw in this paragraph is not the writing style, but the perspective.
Who is the narrator? Someone who sleeps during the day, is active at night, and doesn’t even know the name of his assistant . Just in the previous sentence, he admits: "She knows the names of those two girls. I don’t." In the next moment, he suddenly possesses the history of these girls' craft progressions and internal discussions.
Technically speaking
The model optimizes for "detail richness," not for optimizing "the narrator's knowledge boundaries." A first-person limited perspective is a hard constraint that must be maintained throughout; human authors rely on common sense to uphold it, while models cannot—they only know that this position requires more texture, so they fill in more texture. Any paragraph where the narrator could not possibly know yet is written in extreme detail should raise questions.
Six | Five supporting characters, the same person
I looked at that shadow and thought, who is this person, how much has he walked today, how many houses has he seen, and does he know who this house is for.
Why AI?
A three-part rhetorical question. Moreover, the emotional logic does not hold— a man who is being asked to pay fifty million and is hanging on a marriage would not empathize with an assistant shooting a house video at this moment.
Yet the thematic logic holds. There are five such close-ups of supporting characters in the entire text: a Mexican woman cleaning a vacant room, two girls applying tape, an assistant filming, a surrogate mother whose pictures in the file have never been opened, and a child holding popcorn being dissuaded. The theme of these five segments is completely consistent— doing unknown labor for someone who is not present.
Technically speaking
This is theme-driven writing, not emotion-driven writing. When a person writes about their own breakup, the distribution of attention is imbalanced, selfish, and all over the place; the model’s attention is naturally uniform, spreading the theme evenly throughout the piece. The five supporting characters are written equally well, equally briefly, and equally restrained—this uniformity itself is a signature.
Seven | A sentence layering two templates
I have never rejected her; it is not generosity, it is fear. Fear of the answer in the next second.
Why AI?
The first half is the "not X, but Y" negation-correction framework, and the second half is a chaining: the last word of the previous sentence directly becomes the first word of the next.
Technically speaking
Both are high-frequency outputs of a large model. The negation-correction format is often overused because it can create a "sense of profundity" at extremely low cost: it first negates a common explanation, then provides a slightly unexpected word, making it read like an insight. Chaining is more mechanical—the attention weight of the freshly generated word is peaking, and reusing it is the easiest choice.
"Not X, but Y" appears at least three times in the text. Humans would not become so addicted to the same sentence structure over six thousand words.
Eight | A word that is out of place
What chills my bones is not that it is smart, but that Claude shows no hesitation at all. / It does not hold back.
Why AI?
Another instance of negation-correction, followed by a three-word clause that stands alone to conclude.
Another word feels out of place: "chills my bones". This is a formal big word that is completely out of sync with the colloquial restraint of phrases like "I said good," "that’s fine," and "none of your business." Such high-level words wedged in during polishing are often isolated formal terms — surrounded by plain speech, suddenly elevated by a notch.
Nine | Measuring heartbeat
The heartbeat is very steady. I counted again, still very steady. … Like a doctor conducting a check-up on someone else. / I did not cry. I have been waiting for myself to cry.
Why AI?
The use of the body’s nonreactions to convey the enormity of emotion is the model's highest hit rate trick for writing "restrained sadness", a template-level expression. "I did not cry. I have been waiting for myself to cry." is the standard sentence structure for negation plus waiting. "Like a doctor conducting a check-up on someone else" is another signature simile.
Technically speaking
These types of sentences have extremely high density in the training corpus—they are the "shortest path to a literary feel". When prompts include requirements like "restraint" and "do not state emotions directly," they are almost inevitably summoned back. They are well-written, but the problem is that they are standard to a degree that they could directly serve as teaching examples.
Ten | The staircase at the end
A million people surge in like a tidal wave, and as the fireworks end, a million people retreat. / A million people. / It is always like this every year. / Hong Kong at three o'clock in the morning, / nothing has happened.
Why AI?
Each line is shorter than the previous one, gradually converging; "a million people" stands alone to serve as lexical retrieval; finally, it precisely lands back on the thematic words buried at the beginning. All three things are accomplished at once.
By the way, notice "surge in like a tidal wave"— a cliché metaphor that is clearly of a lower standard compared to the cold, stark portrayals earlier. Areas with uneven levels of quality usually indicate seams where splicing occurs.
Technically speaking
Stair-step endings are the strongest among concluding templates. The model knows that the "ending" position requires deceleration, recovery, and leaving blank spaces, so it executes all three together. Human endings often lack completion, missing one or two elements, or simply fail to bring it to a halt.
Additionally, the phrase "I said good" appears repeatedly throughout the text, acting as a metronome. Effective, but mechanical. Once a model establishes a refrain, the easiest response for every subsequent dialogue is that, and the more it is used, the stronger it becomes.
04
The harder evidence is what is absent
The ten instances above belong to "the things that are added." There is another category—the things that are missing, which are often even more telling.
There are almost no conjunctions in six thousand words. No "however," no "therefore," no "at the same time." This is not how natural writing appears; this is "the instruction to avoid conjunctions has been fully executed."
There are almost no exclamations in six thousand words. No modal particles, no vulgar language, not a single self-justification, not a pitch off-topic, and nothing repeated twice.
A man who hasn’t slept for two consecutive nights and is just being asked for fifty million dollars does not write six thousand words without swearing once or defending himself in any way— this is impossible.
In moments of heightened emotion, people lose control of language: they repeat, break sentences, suddenly curse, write entire paragraphs that are completely useless, and contradict themselves. These "useless bits" are the carriers of the human touch. The model does not produce useless talk because useless talk has never been the optimal solution probabilistically.
05
This does not resemble Sun Yuchen, any version of Sun Yuchen
Let’s put out three samples of his public corpus.
The first one, from 2007. At seventeen, he won the first prize of the 9th New Concept Essay Contest, earning admission qualifications for Peking University. What was the writing style of that generation of new concepts? Lyrical, rhetoric-intensive, with shining gems, with "I" always at the forefront.
The second one, around 2008. He published "A Proof Argument" in "Mengya", where the core is a declaration—that one year in high school can make up for any regrets. The conclusion comes first, the entire piece argues, and he leaves a contact address at the end inviting others to validate it. It is said that he received nearly ten thousand letters within two years. This is the standard argumentative persona: offensive, persuasive, and recruiting.
The third one, the past ten years. Exclamation marks, all caps, sweepstakes, trending topics, big words, emojis. The writing style of the king of crypto marketing has remained unchanged for ten years.
These three samples span nineteen years, and their commonality is very clear: outgoing, offensive, strongly self-centered, never shy about voice.
What is "My Girlfriend Jing Tian"? Introverted, restrained, neutral, portraying oneself as a limited observer who can manipulate everything but decides nothing.
A person can certainly write something that does not resemble themselves under special circumstances. But a high-frequency writer's punctuation habits, modal particle habits, and rhetorical impulses are hard to clear out at the same time on the same night. You can imitate someone else’s wording, but it’s hard to imitate someone else’s breath.
The breath of this article is not his.
The only thing that hasn’t changed is the purpose—it is still meant to persuade you. It’s just that this time, the method of persuasion is rented.
06
What he wrote himself
Looking at it in comparison, it is clearer. These parts I believe are basically original:
Adding friends on Renren, spending forty minutes deleting and rewriting the message "hello," only to end up not sending it.
A one hundred fifty yuan down jacket, hesitated for three days.
"Call me mom", and subsequently in the narration, the address changes from "Jing Tian" to "mom", until it slides back to "Jing Tian" the moment the call ends.
Cannot mention that athlete.
Tearing away the ranking, the couple's dialogue—rude tone, heavy with information, clearly not on the same wavelength as the surrounding literary tone.
Jing Guoqing, Tian Xin’ai two names and two accounts; one card limited to one million per day, so he transferred five; hesitated for a moment when moving to the third card.
The pattern is so clear that it could almost be used as a ruler: the more specific, more unflattering, less dignified the place, the more it resembles what a person writes; the prettier, more nuanced, and more suitable for screenshots and shares, the more it resembles machine polishing.
This is because unflattering details are hard to invent; they can only be remembered. While beautiful sentences can be generated.
07
He wrote the tools into the work
There is a whole section in the middle of the article: he said he used Claude Code to run through cash assets, then asked it a question, and the answer was not to give that fifty million; he followed up asking "does she no longer love me?" AI's response was essentially that he does not care about love nor understand it, but that he cannot give it.
Two things.
First, this "Claude" is not a tool, it is a character. A real programming assistant would not respond "Does she still love me?" This dialogue is written as a literary figure, not a log.
Second, and more importantly: someone who has not interacted with AI while writing would not think to make AI a key supporting character in the plot, let alone write dialogue for it and describe it as "it doesn’t hold back."
On the night the article was published, some people on tech forums were already saying "this feels just right" and "like it was polished by Claude." The threshold to judge this is actually quite low for those who use these tools daily.
Writing tools into the work serves as an advertisement and is also a semi-public signature.
08
Two pieces of corroborative evidence
One, he turned this piece into a LaTeX project and uploaded it to GitHub, named the repository my-girlfriend-jingtian-latex, with a PDF. Staying up late to manually format a long breakup article in LaTeX—this is not a writing process, this is an engineering process.
Two, within days after the article went viral, several repositories popped up on GitHub specifically replicating this writing style, breaking it down into several executable rules: extremely short sentence rhythm, dense numeric anchors, emotions never explicitly stated, and philosophical closure, referencing Hemingway's iceberg theory as a technique.
When a writing style can be cleanly broken down into several prompt words, it usually indicates that it originated as a series of prompt words. Genuine personal styles cannot be separated this neatly—because personal styles contain large amounts of unreasonable elements, while prompt words can only encapsulate reasonable parts.
09
In conclusion
Sun Yuchen provided all the materials, all the emotions, all the decisions; AI provided the sentences.
Using tools is not a scandal. Those painful parts in the six thousand words are indeed painful.
What is truly worth noting is something else. This may be the first time on the Chinese internet that a top billionaire's personal emotional statement also serves as litigation material, a branding campaign, a GitHub repository, and a piece of AI-generated text. Four layers of content seamlessly sewn together and presented quite aesthetically.
In the future, such content will only increase. Recognizing stitching is becoming a fundamental skill.
Verification notes
· Confirming facts: On August 27, Sun Yuchen published "My Girlfriend Jing Tian" on X, with both the beginning and end marked "This article is purely fictional"; attorney Zhang Qihua confirmed the lawsuit against Jing Tian and her parents, targeting over thirty million yuan, which the court has filed; Jing Tian's side raised objections to jurisdiction, and the case has not yet entered substantive hearings; Jing Tian's studio issued a statement denying the related rumors.
· Confirming facts: Sun Yuchen won the first prize of the 9th New Concept Essay Contest in 2007, later attended Peking University and published "A Proof Argument" in "Mengya" during his first year.
· Single source: Claims about writing taking over ten hours, two nights, and being in Bhutan come from statements made by Sun Yuchen himself during interviews without third-party verification.
· Author’s estimation: The weight of fifty million dollars in cash calculated based on the author's estimations, as the original text did not specify denominations.
· Author’s inference: All segments identified as AI-generated or polished are based on inferences regarding the text’s style and generation mechanism, and do not constitute factual accusations. Sun Yuchen has not provided any clarification on this matter.
· Quotes from the original text are short segments used for textual analysis, and the full text can be retrieved independently.
· All plots involving Jing Tian in the text come from Sun Yuchen's unilateral narration, which has been denied by Jing Tian's side; this article does not make any judgments about its authenticity.
—— Kea
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
