AI正在吃掉自己的粮食，而且粮食越来越馊 (English)

Generated: 2026-06-22 04:33:59

---

Friend, have you ever noticed something deeply unsettling?

Come, let's do an experiment. Open Zhihu, search for a popular question, then scroll past the tenth page. Look at those answers—standardized wording, rigorous logic, neat paragraphs, reading like a textbook-perfect piece. But don't you get a feeling? They lack soul. Like they were stamped from the same mold.

You guessed it. They were most likely written by AI.

What's even scarier? When the next generation of AI is trained, it will devour this kind of content as "human knowledge." Then its output will become even more like these AI-written pieces. And then the next generation of AI eats that. An infinite nesting doll.

What do I call this? — "AI's fecal cycle."

I'm not just crying wolf. The data from January 2025 is right there: over 51.72% of new content on the internet is AI-generated, and this number is skyrocketing.

Someone ran an experiment, having AI train on its own output, generation after generation. Guess what happened?

First generation, not bad, pretty much like a normal person. Third generation, started to get a bit dull, repeating itself. Fifth generation, noticeably dumber, forgot all the niche knowledge. By the ninth generation—"gone mad." You ask it about medieval architecture, and it rambles on about colorful rabbits. Why? Because medieval manuscripts are indeed filled with colorful rabbits; the model memorized it by rote, thinking the two were the same thing.

This isn't a joke. This is real "model collapse."

I've fallen into this pit myself. Last year, while building a chatbot for a niche field, I got lazy and used a ton of data scraped from the web for training. What happened? The model's understanding of technical terms drifted further and further off, until it could explain "TCP three-way handshake" as "shaking hands three times to establish a connection." Only later did I realize those "professional articles" were all AI-written, filled with plausible-sounding but wrong concepts.

See, this is what I mean—AI is eating its own food, and the food is getting more and more rotten.

Speaking of which, I want to talk about two issues with you.

The first issue comes from OpenAI's Ilya Sutskever, who said at the NeurIPS conference at the end of 2024: "Data is the fossil fuel of AI." The logic is clear: AI got stronger in the past few years by relying on more parameters × more data × more computing power. Computing power is still increasing, but the data leg has hit a wall—there's only so much high-quality text on the internet. All the books, papers, news, blogs, and code combined form a fixed pool. The big companies have basically scraped everything they can. There's no "second internet."

He's talking about a stock problem: the total amount of human knowledge is limited and has been mostly used up.

The second issue is mine. I'm talking about a flow problem: in the future, the amount of newly generated high-quality content will keep decreasing because the spread of AI is eroding the motivation to create and the quality of content.

These two issues are two sides of the same coin. If you look at the scale, his problem is more urgent—decades of accumulated text might be trillions of tokens, with newly created high-quality original content making up a tiny fraction each year. But if you look long-term, my problem is more dangerous—he describes a static ceiling, while I describe a dynamic deterioration: not only is there no new fuel, but even the existing oil fields are being polluted.

Think about it. When was the last time you seriously wrote a long post sharing your experience on Zhihu? When was the last time you asked an AI a question?

Between these two questions lies a fatal connection.

In the past, when I encountered a tricky programming problem, I'd post on Stack Overflow and wait hours or even days for some expert to enlighten me. Now I ask AI directly and get an answer in 30 seconds. In the past, after reading a good book, I'd write a review on Douban, enjoying the exchange with other readers. Now AI can write a better review than me and summarize the whole book. In the past, after traveling somewhere, I'd post a guide on Xiaohongshu, collecting likes and saves. Now I let AI plan the itinerary, and it can even dig up niche spots locals don't know about.

So here's the question: if everyone stops sharing, where will future AI find new data?

This is a terrifying closed loop: AI's growth depends on human sharing in public spaces, but the spread of AI is destroying the motivation to share. AI is "eating" its own future.

After figuring this out, I suddenly understood why the New York Times is suing OpenAI, why Disney is suing OpenAI.

My first reaction used to be "old forces hindering innovation." Now, looking back, these lawsuits are actually helping the AI industry solve a problem it can't solve on its own.

Without legal pressure, AI companies have no incentive to pay voluntarily. If there are no consequences for using free content to train models, the rational choice is to keep taking it for free. OpenAI's licensing agreement with the Associated Press and its negotiations with various publishers mostly accelerated after the New York Times lawsuit.

Moreover, these lawsuits are establishing rules. The legal boundaries of AI using copyrighted content are currently very blurry, with no precedents. Once these cases are settled, they might form a pricing system for "data usage fees"—just like music copyrights or image copyrights.

Some might say, "Isn't this just a disguised monopoly? Big companies buy data with money, small companies can't afford it, and in the end, only a few giants can play."

That's true. But the problem is, relying entirely on the market won't work either. An independent blogger has no idea which models have used their articles and has no bargaining power. A pure market outcome would be big institutions signing high-price agreements while long-tail creators get nothing.

This problem might not have a "correct institutional design." It's essentially the same as all public goods dilemmas in human history—once knowledge is produced, it can't be exclusively controlled in its spread; forcibly controlling it stifles efficiency, and not controlling it leaves producers unrewarded.

Based on this judgment, I've come to a straightforward and realistic prediction: In the future, the ability to simply excel at text creation or mechanical coding will no longer be scarce. What will truly be precious are people who can think independently and engage in deep, impactful exchanges.

From now on, humans themselves will become the most essential raw nourishment for the AI era.

Looking back at the process of intellectual collision, it's easy to understand: without wild ideas, without novel cross-disciplinary concepts, AI loses the source of creation and can only mechanically churn out uniform content. Humans are a "mine" of infinite ideas, while AI is just a tool for mining, organizing, and implementing. Without source ideas, the tool is useless.

Those with clear logic, broad thinking, the courage to break through established cognition, and the ability to connect seemingly unrelated things into new perspectives will become a scarce resource for the entire era.

To be honest, I don't have a perfect answer. But there are a few directions worth trying:

First, add a "Human Original" label to your content. Just like organic food certification, we might need a "Human Original" certification system in the future. This protects the value of your creation and provides "clean" training data for AI.

Second, don't let AI think for you. Using AI as an assistant is fine, but don't let it make decisions for you. My current habit is: first use AI to gather information and organize a framework, then turn off AI and write it myself. After writing, I let AI check for logical flaws. This way, I use the tool while maintaining independent thinking.

Third, focus on areas where AI struggles. AI's accuracy on causal reasoning questions is still very low, lower than an elementary school student's. It lacks "metacognition"—it doesn't know what it doesn't know. It has no emotions, desires, or intuition, and can't experience sensory pleasure. These areas are precisely human opportunities.

Finally, let me say this: AI won't eliminate humans, but it will eliminate those who give up thinking. Every genuine original post on the internet, every piece not generated by AI, is becoming a scarce resource. And scarcity means value.

So, friend, the next time you're about to click the "generate" button, pause. Ask yourself: Is this really something worth handing over to AI?

Because in this age of AI saturation, human originality is the rarest, most expensive luxury.