The Reason AI Writing Tastes Like Nothing

AI flavor: where it comes from and where it goes

At 1 a.m., you paste the AI draft into the document. The third paragraph begins with those six characters again: “not... but...” You delete it. Refresh. It puts on a new disguise: “It’s not so much... as...” You set the water glass on the corner of the desk. A ring of tea stain sits at the bottom. You remember the meeting earlier that day. Your boss said, “We need to form a closed loop.” You nodded. Your stomach turned sour.

AI flavor has a recipe.

The big pot

Pretraining is a big pot. The internet, books, papers, news, official documents, ads, forums, WeChat public accounts, speeches. All dumped in. Turn the heat up. Boil. What comes out is the lowest common denominator. A distinct personality is like salt thrown into the sea. You can’t taste it. You take a sip. Not salty, not bland. Drinkable, but you don’t want another sip.

Canteen big-pot food is like this. It isn’t delicious. It isn’t terrible. It’s safe. The default AI style is canteen big-pot food: leaves, minced meat, starch thickener, no wok hei.

Why it always loves “not... but...”

This sentence pattern is the universal wrench of public writing. It first loosens your original thought, then tightens a new one. Speeches use it. Ads use it. WeChat public accounts use it. Leaders’ speeches use it. It creates the thrill of “you thought wrong before.” Any topic can fit. Write about technology, life, love, weight loss, international affairs. All can fit.

The model has read this continuation billions of times. It doesn’t know the sentence is worn out. It only knows that in this position, these six characters have high probability. Turn the temperature low, and it loves this main road even more.

Go look at viral headlines, meeting minutes, brand manifestos. They all teach the model the same thing: first negate, then affirm, and you will look profound. The model learned. It learned too well.

“Push, hard, steady” are like instant noodle seasoning

“Push” can form: push forward, promote, spread, implement. “Hard” can form: hardcore, hard power, hard tech. “Steady” can form: steady, stable, seek progress amid stability.

These words share one thing: they are never wrong anywhere. They have no image anywhere. Like instant noodle seasoning packets. Sprinkle it on braised beef noodles, mushroom chicken noodles, seafood noodles. After sprinkling, it all tastes the same. You can’t taste beef. You can’t taste chicken.

When the model is uncertain, it chooses safe words. Safe words don’t offend, don’t take sides, aren’t specific. Being specific can be wrong. Being vague cannot.

The annotator’s finger

In later training, human annotators look at two answers and click one. They usually click: clear structure, balanced tone, no insults, has a summary. The model remembers this finger. It learns: first define, then list, then compare, finally add a summary. This gets clicked.

So AI boilerplate takes shape. It is like neatly written meeting minutes. Every sentence is right. Every sentence is useless.

Temperature and sampling

During generation, the model chooses words at forks. Low temperature: it takes the main road. The main road is wide, crowded, safe. High temperature: it takes a side path. It may fall. It may also see wildflowers. Default settings tend to be cold. So “not... but...” appears again and again. “Push, hard, steady” appears again and again.

If you only give a topic, you are telling it: walk anywhere. It goes back to the main road.

The prism

AI has personality. Conditions suppress it. When you give it a prompt, you give it a prism.

You say: Explain quantum mechanics in the voice of a foul-mouthed guy from Northeast China. It immediately changes person:

Listen up! Quantum mechanics, this thing. The microscopic world doesn’t play fair! Where is an electron? It’s everywhere and nowhere! When you look, it behaves. When you don’t, it runs wild!

You say: Write a resignation letter in the voice of Lin Daiyu. It changes again:

I had thought to find a place to settle. Who knew I would be living under another’s roof. Ground down by performance reviews, I no longer have the strength to wear this work badge.

Those styles did not disappear. They were averaged, canceled out, hidden in the conditional distribution. Shine a prompt on them, and the colors come out.

How to break it when writing

Don’t just give a topic. Give seven things.

Role. Who is speaking. Sharp-tongued critic, laid-off worker, owner of a small shop in a county town, Lin Daiyu, Northeastern big brother. Role determines vocabulary.

Scene. To whom. Late-night WeChat Moments, company group chat, email to investors, letter to an ex. Scene determines measure.

Taboos. Banned words. List “not... but...”, “empower,” “lever,” “closed loop,” “underlying logic,” “hardcore,” “steady,” “promote,” “build,” “reshape,” “promising future.” Ban summary and elevation. Ban parallelism in every paragraph. The more specific the taboo, the less the AI dares to slide back.

Sample. Paste a passage you like. Don’t just say “like Wang Xiaobo.” Paste 100 words. Let it imitate sentence length, rhythm, punctuation, metaphor.

Details. Give time, place, numbers, actions, dialogue, objects. If you give “3 p.m., the convenience store has only one fish ball left in the oden,” it won’t just write “life is hard.”

Conflict. Write two people arguing, a decision regretted, an interview failed, a lie exposed. Conflict forces concrete language.

Rhythm. Short sentences mostly, occasional long ones. No more than three sentences per paragraph. Don’t summarize every paragraph. End on an action or an image.

Self-check decision tree

After each emotional shift, stop. Ask three questions.

Is the skeleton there? Delete this paragraph. Can the text before and after connect cleanly? If yes, this paragraph is redundant. Rewrite it. If no, pass.

Is the flesh unique? Replace the names with Zhang San and Li Si. Does this paragraph still hold? If yes, it’s too generic. Replace it. If no, pass.

Is it honest? Is this emotion said, or seen? If said, go back to the action level. If seen, pass.

Then search words. Search “truly,” “really,” “it must be said,” “worth pondering,” “marks.” If any appears, delete the whole sentence. Search “he felt,” “she realized,” “he understood.” If any appears, replace it with what he did. After an emotional high point, immediately follow with an extremely ordinary detail. Stomach medicine, pothos, keyboard cover. Press the emotion back to the ground.

Exercise

Take a topic. Write three versions.

First version, default. Let AI write freely. Observe what boilerplate it uses.

Second version, role. Set role, scene, emotion. Ban ten high-frequency words.

Third version, detail. Give no opinion. Give only five details: a time, a place, a number, an action, a line of dialogue. Let AI grow an article from the details.

Compare. The first version looks like AI. The second has human flavor. The third looks like life.

Back to 1 a.m.

You delete “not... but...” You write: The red light on the kettle base is on. You press the switch. The water boils. You drop the tea bag into the cup. The tea bag sinks, then floats up. You sit back in the chair. You change the prompt to: In the tone of a sharp-tongued tech critic, ban those ten words, write 200 words. The cursor blinks once. On the screen appears the first sentence: At 1 a.m., you are still deleting “not... but...”

You smile. This time you don’t delete it.