The “Cooking Without Rice” of AI Writing: Why You Can’t Write Good Articles with AI?

Open Claude or ChatGPT in your browser and type a smooth prompt into the dialogue box:

“Please write an analysis of the business model of the cross-border e-commerce giant Temu. Requirements: clear logic, multi-angle analysis, around 3000 words.”

Press Enter. Within a mere 30 seconds, the cursor on the screen starts scrolling down at an extremely high and fluid speed, like an tireless typist frantically hitting the keys.

In less than a minute, an article is perfectly presented before your eyes.

It has beautiful Markdown formatting, with symmetrical and neat subheadings: “I. Background of Temu’s Rise; II. In-depth Analysis of Core Business Model; III. Supply Chain Dividends of Ultimate Cost-Performance; IV. Global Compliance Risks Faced…”

Every paragraph is written flawlessly, with not even a single punctuation error to be found. It talks about how the full-service model shortens the supply chain, discusses the reshaping of domestic overflow production capacity, and mentions the uncertainties brought by geopolitical factors.

You read it word for word.

After finishing, a strange emptiness suddenly arises in your mind. It’s a physical reaction similar to the slight stomach cramps after drinking a tasteless and calorie-free meal replacement powder.

The article seems to have everything, yet it also seems to say nothing. It’s like a bowl of braised pork made of plastic, reflecting a shiny red luster under the light, with an even more regular texture than real meat, but it has no warmth, no rich meaty aroma bursting in your mouth. When you swallow it, all you feel is coldness and hollowness.

This is the unique “plastic feel” of the AI era.

You begin to feel a sense of void. As a writer, you think: Since AI can write an article with such a complete structure, comprehensive arguments, and appropriate rhetoric in 30 seconds, what is the meaning of my writing? Is all my hard work of thinking and typing just to compete in a losing race against a probabilistic machine that can spit out hundreds or thousands of tokens per second?

This disheartening question isn’t about insufficient efficiency or not smart enough large language models. The problem lies in a deeper place that most people never even notice: when we expect large language models to handle all content production with one click, we are actually asking it to do “cooking without rice.”

The Silent Labor Beneath the Waterline

Before a single word is typed on a typewriter, with a pen, or on a keyboard, most of the work of writing is actually already done.

Looking back at the history of deep human creation over hundreds of years, true writers all know that writing is a classic “iceberg game.” The part exposed above the waterline—the text that readers can see—often accounts for only 10% to 20% of the total workload. The remaining 80% is the massive, silent, and extremely torturous research work at the bottom of the iceberg.

In the age of libraries and archives, non-fiction writers and investigative journalists spent most of their lives among paper scraps and dust. They needed to write letters to contact experts, take trains to other cities to review local gazettes, and strain their eyes at microfilm readers, just to verify a date or find a key detail that could support the core narrative.

At that time, “finding material” was the fundamental skill of a craftsman. If the material wasn’t solid enough, no matter how flashy the writing, the output would be flimsy paper.

Around 2000, search engines changed the rules. The explosion of the internet compressed the vast paper library into the input box of a browser.

Finding information became democratized. You no longer needed to run your legs off; just enter keywords, and hundreds or thousands of links would appear before you. In this era, “collection” became active discovery and filtering. Although you no longer needed physical effort, you still had to click different links in your mind, cross-reference viewpoints, judge the credibility of information, and through this process, painstakingly build your own understanding framework of the issue.

In the age of Google and Wikipedia, writers still experienced this filtering pain. And it was this pain that served as cognitive friction, forcing writers to think: Why doesn’t this data match that one? Why are the accounts of these two parties completely opposite?

However, the sudden explosion of large language models (LLMs) at the end of 2022 promised, in an extremely shocking way, to flatten all this pain.

The interface of LLMs is a simple, clean dialogue box. It tells you: You don’t need to open hundreds of links anymore, nor do you need to endure the tedium of reading long PDF papers, and you don’t need to worry about complex logical transitions. Give me your topic, and I will directly give you a complete article, even with subheadings already prepared.

This promise was so sweet that people almost instantly dropped all their “defenses.”

People began to believe that the bottom of the iceberg was no longer important. AI had already packed all the world’s knowledge into its neural network; we just needed to give it a few gentle commands, and the tip of the iceberg floating above the water would materialize out of thin air.

In this frenzy of “one-click generation,” the identity of the writer was quietly changing. We degenerated from explorers and gold miners in the information world to mere issuers of commands in the dialogue box. We no longer experienced the pain of sifting through garbage to find gold, and thus naturally lost the “epiphany” moment of discovering gold in the sand and mud.

This is the source of the weightlessness. When the step of information gathering is compressed to near zero, writing loses its gravity. It becomes light and fluffy, like a bubble blown in a vacuum, although colorful, it will disperse at the slightest breeze.

We thought we had skipped the most tedious and exhausting chores, achieving ultimate efficiency. But we never expected that the collection and digestion process we threw away as garbage is actually the core quality of the entire iceberg.

Mean Reversion: Statistics Has No Miracles

The assumption that because the large model has read the whole world, it can write the most profound insights sounds extremely logical, but encounters the most awkward failure in reality.

The truth is the opposite: precisely because the large model has read the entire human internet, when you don’t provide it with specific materials, what it writes is destined to be the most mediocre.

We can analyze this paradox from the principle of statistics. The essence of a Large Language Model (LLM) is a “probabilistic continuation machine.” When you input a prompt, the action it performs is to calculate and output the most reasonable next word (token) based on a massive amount of pre-training data. This process is called “Maximum Likelihood Estimation” (MLE).

What is “most reasonable”? Most reasonable means highest probability, that is, it best conforms to the “average level” of internet text.

When you say to AI, “Write a deep analysis of Temu’s business model,” and don’t provide any first-hand, unique real-world data or detailed cases, you are actually pushing it into a completely weightless state of free improvisation. At this point, the large model, in order to complete the task, can only call on its parametric memory—the statistical average of those trillions of tokens it swallowed during pre-training.

It will immediately slide towards that safest, most written-about word combination: “full-service model,” “ultimate cost-performance,” “supply chain dividends.” These words are not wrong, and the logic is watertight. But they contain no new information. They are the “greatest common divisor” on the internet regarding this topic.

Claude Shannon, the founder of information theory, once gave a very sharp definition of information: Information is a measure of surprise. The amount of information a sentence conveys is inversely proportional to its probability. If you say something that everyone completely expects, with a probability of 100%, the amount of information you convey is actually zero.

AI writing from scratch is a typical “zero surprise” writing. Its sentences are extremely smooth, the structure extremely symmetrical, but after reading it, it’s like taking a breath of air, nothing stays. There are no miracles in statistics; the convergence of probability will ruthlessly pull every “free improvisation” towards a mediocre average.

The consequence of this “mean reversion” is not just reader aversion, but also platform cleansing.

In March 2024, Google’s search engine launched an unprecedented quality crusade (March 2024 Core Update). The purpose of this update was clear: to cleanse those websites that relied on AI to mass-produce, out of thin air, low-quality junk content (often called AI Slop).

The intensity of this purge shocked the entire digital marketing industry. According to tracking research by the content detection agency Originality.ai on penalized sites, Google directly removed about 2% of websites (Deindexed) from the search index in this update, rendering them completely invisible on the internet.

The correlation shown by the data is extremely brutal. In Nuttall’s analysis of hundreds of completely deindexed sites, 100% of the affected websites showed obvious characteristics of AI-generated content. Among them, 50% of the websites had 90% to 100% of their pages mass-generated entirely by AI. These websites fabricated hundreds or thousands of “logically coherent, structurally complete” articles from thin air every day, trying to attract search traffic.

As a result, these completely deindexed websites directly lost approximately $446,552 per month in display advertising revenue.

This is not only the bankruptcy of content farms but also the ultimate verdict on the illusion of “one-click generation of deep articles.”

What the large model demonstrates are divine skills in reasoning and compilation; it cannot conjure up a completely new and accurate physical world out of thin air. Abandoning the search for solid, first-hand facts and expecting AI to produce new ideas based on its old knowledge from the neural network is as absurd as trying to squeeze water from air.

The gravity of the average is so strong that only by using high-entropy unique materials as “anchors” can we drag the large model out of the void of statistical averages.

Sending Fresh Seafood to a Master Chef

If you lock a Michelin chef in an empty, bare-bones room and only give him a few packets of “Master Kong instant noodle seasoning powder,” no matter how superb his culinary skills, he can only give you a bowl of industrial-tasting salty soup.

Giving requests to AI empty-handed is like issuing orders to this master chef in the bare-bones room. The large model is that chef with exquisite skills, and the adjectives like “requires depth, rigorous logic, multi-angle analysis” you write in your prompt are just urging him to make the soup slightly more uniform.

For him to create an astonishing state banquet, you must personally go to the market and throw the dirt-covered king crabs, dew-kissed matsutake mushrooms, and freshly slaughtered meat onto his chopping board.

These dirt-covered, dew-kissed raw materials are the search and collection we must undertake before writing.

Why does inputting specific materials lead to a qualitative change in AI’s generation quality? This involves the Attention Mechanism and In-Context Learning (ICL) principles of large language models.

When a large language model without context input is generating text, its attention pointer is drifting in a boundless cloud of prior knowledge probability. It unconsciously grabs the most common, highest-frequency patterns.

However, once you feed a large amount of specific, unprocessed facts, data, and cases into the prompt, this spares it from the task it is least good at—searching and fabricating facts from scratch in its mind. It can then concentrate all its computational resources (Attention Heads) on deploying its true talent: building a rigorous and unexpected logical network between existing, solid coordinate points.

In this process, the “heterogeneity” and “multidimensionality” of the input materials are crucial. If your materials only contain one voice, one perspective, the large model’s logical circuits will still tend to become flat.

This is also why, in the workflow of deep writing, collecting materials cannot be a random “grab a handful of sand,” but must be a set of extremely tense, multi-dimensional collection strategies:

Hook Scenario: A specific, warm, real moment. For example, “a technician at 2 AM staring blankly at a series of error codes on the terminal.” This can forcibly pull AI into the scene, preventing it from using empty openings like “With the development of technology, XXX has become deeply ingrained.”
Historical Context: The depth of time. Collect who proposed this technology or event 10, 20, or even half a century ago, and how it was solved at the time. This lays a three-dimensional foundation for the article, avoiding the AI-specific characteristic of only seeing the present without a past.
Core Data: A quantitative coordinate system. Don’t just collect numbers; collect reference points. For example, instead of letting AI write “This model is very large and consumes a lot of electricity,” feed it specific numbers: “This 300-megawatt data center consumes electricity roughly equivalent to the daily consumption of 300,000 ordinary households.” Once numbers have a reference point, they gain weight.
Technical Details: Deconstructing the underlying mechanism. This requires collecting original papers, development documents, and using specific non-technical analogies for translation (for example, comparing “TileRT” to “a factory switching from batch processing to a continuous assembly line”). This prevents AI from slacking off and forces it to write with real technical substance.
Industrial Cases: The friction of real-world deployment. The metrics in papers are always perfect, but the real-world ledgers and engineering blogs of the industry are full of blood. Collecting these conflicts fills the article with a sense of “real people.”
Opposing Views: A reverse thrust. Collect dissenting opinions, criticisms, or discussions of different technical routes (such as complaints on Reddit or Hacker News). This is the best antidote to prevent AI from falling into one-sided narrative and mindless praise.

When the solid facts from these six dimensions are placed on the chopping board, the probability distribution space before the AI is forcibly distorted. The Attention Mechanism begins to weave logic between these hard, fixed coordinates like “300 megawatts,” “20% academic cheating rate,” “Google’s 2% penalty data.”

At this point, the master chef starts to truly perform his divine skills: he can use his astonishing ability to structure and assemble these materials, use the most clever sentence transitions to connect them, and even effortlessly conjure up an unexpected and stunning analogy.

You let the large model do what it should do—compile and reason—while you do what only humans can do—search for and transport unique facts in the chaotic, complex reality.

This is how high-entropy material input completely reshapes the AI generation mechanism.

From Word Craftsman to Information Curator

In this era, the top AI writers and the most foolish AI writers are heading down two completely opposite paths.

The first path leads to the desert of content farms. On this path, users treat AI as an automatic typewriter. They input prompts like “Write an article about the future development of new energy vehicles” every day on an assembly line, and then copy and paste the tens of thousands of words generated by the model verbatim.

Before 2024, these AI junk articles that relied on keyword stuffing and superficially symmetrical logic could still get a share of the pie from search engines via SEO algorithms. But after Google’s algorithmic baptism, these websites are turning into digital ruins at an alarming speed. They are being demoted sitewide or completely deindexed (K-stationed), proving that cooking without rice, without unique fact input, has no way of surviving under the joint strangulation of humans and algorithms.

The second path, however, leads to the deep space of high-entropy curators.

Excellent practitioners on this path have long abandoned the naive fantasy of letting AI write articles from scratch. On the contrary, they place extreme importance on input quality, even dedicating 80% of their daily energy to information Curation.

If we analyze the workflow of those technical writers who have written articles with millions of views and extremely hardcore logic, you will discover a startling fact: in their entire writing process, AI is prohibited from participating in any “fact discovery” steps.

They would never ask AI in the dialogue box a common-sense question like “How does Temu implement its full-service model?” They would rather spend three days going through Pinduoduo’s annual SEC financial reports, searching Hacker News for complaints from overseas merchants about Pinduoduo’s logistics fees, lurking on Reddit to see consumers’ real experiences with Pinduoduo’s delivery times, or even downloading a dozen academic papers on cross-border supply chain logistics efficiency.

They would build a highly structured, highly credible library of first-hand facts in their local software. In this library, every piece of material has a precise source URL, accurate data points, records of opposing viewpoints, and credibility ratings.

This fact library is the high-gravity field they build for the AI.

Only after this step is completed do they activate the large model. They feed this meticulously curated fact library to the AI, with strict constraints like this:

“Please draft the framework for Section 3.2 based on the fact library I provided. Every conclusion you write must be supported by a corresponding data point in my fact library; if the fact library doesn’t have it, you are absolutely forbidden from making it up—it’s better to tell me the material is insufficient.”

In this workflow, these writers have upgraded from traditional “word craftsmen” who type away laboriously to “Information Curators,” “Fact Gatekeepers,” and “Logic Compilers.” In an era where rhetoric is fully commoditized, the premium for discovering facts has been amplified tenfold.

Here, AI is not a ghostwriter, but a rigorous logic compiler. It uses its unparalleled linguistic talent to compile the hard, solid, first-hand facts that the writer painstakingly collected into the most fluent, highly readable, high-quality article.

Comparing the tragic collapse of content farms in 2024 with the sudden rise of deep human-AI collaborative writers in the field of hardcore technical articles, reality offers the clearest message: the battlefield that determines victory or defeat ends long before you open the AI dialogue box, in that dusty, hands-on step of retrieving, chewing, and filtering first-hand facts.

Finding Your Own “Information Friction”

When words are no longer precious, thinking becomes the only luxury.

Large language models have reduced the cost of text generation to near zero, but this doesn’t mean the act of writing has depreciated. On the contrary, it has made the threshold for truly valuable creation unprecedentedly high. Because in the past, you could mask the emptiness of content with exquisite writing and symmetrical paragraphs; but today, AI can write more polished, more gorgeous parallel sentences than you in 10 seconds.

The filter of writing style has shattered, and the hardness of facts is now exposed.

If you want to maintain a writer’s independence and sense of value in the AI era, the most core change you need to make is to re-examine the point of entry for your collaboration with AI.

Don’t expect AI to replace you in colliding with this complex, chaotic real world. AI can only operate in the virtual code world and the pre-trained text space. And you, possessing a living body, can feel real pain, can notice those subtle frictions that defy common sense.

This is your irreplaceable “information friction.”

In practice, this means that the next time you prepare to write a deep article with AI, you can immediately change your habits:

First, close that one-click outline generation prompt template. Stop telling AI, “Help me plan an outline for an article about XXX.” Before writing the outline, force yourself to complete the mindless material collection stage. Read 3 papers with opposing viewpoints, search Twitter for 5 real user complaints, flip through 1 financial report. When you don’t yet have a perfect framework in mind, but you already hold a handful of oddly shaped, even contradictory Fact fragments in your hand, the best creative state has just begun.

Second, treat AI as a stress tester, not an answer repeater. When you have collected these contradictory materials, don’t ask AI how to write it more smoothly, but instead slam the two contradictory details right in its face:

“Look here, this technical paper says this algorithm improves runtime efficiency by 40%, but this tech giant wrote in their engineering blog in 2024 that they had to roll it back in real deployment due to memory overflow. Help me analyze what unnoticed systemic tension might exist here.”

At this point, AI can no longer spit out those average platitudes. It must mobilize its powerful logical reasoning ability to help you compile a path of insight, full of tension and readability, that no one has written before, between the “academically perfect 40%” and the “realistically cruel memory overflow.”

Third, retain the “sense of struggle” in writing. If after finishing an article, you didn’t feel any mental fatigue, nor experienced the process of scratching your head because the materials didn’t match and having to re-search, then it is most likely just a mediocre piece of AI junk content. Moderate difficulty and struggle (Cognitive Desirable Difficulties) are the only signals that your brain is engaged in deep learning and building cognitive connections.

The act of writing itself has never been simply about recording already formed ideas. It is a dynamic craft where we, in the long labor of searching for, transporting, sorting, and piecing together these real-world facts, gradually bring order to our chaotic minds and shape our thinking.

In this era of information overload and homogenization of information, stop being the “command issuer” sitting in an elevator, pointing and gesticulating.

Go into the real market, and find those dirt-covered, fishy-smelling, most real, low-probability facts. Pick them up, stuff them into the attention mechanism of the large model, and then say to it coldly:

“Come, compile these into a true state banquet.”