AI Slop: When the Cost of Content Approaches Zero, There Is No Turning Back

I. AGI: From Euphoria to Reflection

After 2023, the narrative around AI online has taken on a peculiar binary structure.

On one side, there's the AGI euphoria. Every new model release triggers a wave of posts proclaiming a "countdown to the extinction of human professions." OpenAI's launch events have achieved a ceremonial gravity comparable to Apple's keynotes of yesteryear, with onlookers holding their breath for those few benchmark numbers—as if, should the curve bend toward the upper right, the AGI singularity would arrive tomorrow.

On the other side, there are reflections worth noting. Andrej Karpathy recently mentioned on a podcast that he now uses AI-generated articles for initial screening—not to gather information, but to find things that "haven't been written by AI." Gary Marcus has repeatedly discussed how "the average quality of AI-generated content is converging toward the mean."

This phenomenon finally has a name: AI Slop.

Slop, originally meaning "mud, waste, or refuse," now refers to content that is mass-produced by AI—neatly formatted, fluently written, but essentially hollow.

II. Anatomy of AI Slop: The Dumbed-Down Version You've Heard a Thousand Times

Before discussing the essence of AI Slop, you've definitely encountered the following patterns countless times.

"It's not X, but Y"

"Prompt Engineering is not some esoteric art, but a thinking framework." "AGI is not a distant fantasy, but a reality unfolding right now."

This sentence structure conveys zero information. It uses a strawman "wrong view" to highlight a "right view" that says nothing at all.

Strange analogies that explain nothing

"Large language models are like a sponge of human intelligence—they absorb all the knowledge on the internet and squeeze it out when you need it." "The Attention mechanism is like when you're looking for a book in a library, your eyes naturally land on the titles on the spines, rather than scanning every line."

These analogies are technically all wrong, but they "paint a vivid picture" and rack up likes.

Numbers that create urgency, even if fabricated

"ChatGPT gained 1 million users in 5 days; Netflix took 3.5 years—what does this mean?"

What does it mean? The article doesn't know. It just throws the number at you and moves on to the next paragraph.

"The simplest explanation for you" — Doubao style

"Don't be intimidated by those complex concepts, let me give you the simplest explanation..." "Let me tell you in plain language what the Transformer architecture is..."

Then you get an analogy that's harder to understand than the original definition.

The three-part universal structure: What it is → Why it matters → How to do it. No matter the topic, plugging it in yields 500 words.

Transition sentences at the end of every paragraph:

"So, what does this mean for us? Let's continue."

No human being would ever write that sentence in their own article.

III. Why LLMs Inevitably Morph into Clickbait Shape

This is the part I want to focus on, and it's where things get genuinely interesting—LLMs won't surprise you.

The training objective of an LLM is, in essence, to fit human language data. Whatever content is most common on the internet is what LLMs are most inclined to output.

LLM text generation is an autoregressive process, where each token is conditioned on the preceding context. Once the first few paragraphs establish the contextual frame of a "popular science article," every subsequent token is pulled along by that frame. It's like someone picking a song at KTV—once the melody starts, their mouth involuntarily begins to sing along, not because they want to, but because the pattern has been activated.

Clickbait has an extremely strong formal gravity: a direct opening, subheadings, one point per paragraph, and an uplifting conclusion. This structure is incredibly dense on the internet, so in the LLM's "phase space," it forms a very deep potential well. After generating a few sentences, the model almost inevitably slides into it.

So AI Slop isn't a problem of "AI writing ability being insufficient." The essence of the problem is that "AI is optimizing for the wrong objective." It's trained to generate content that humans perceive as good, not content that is actually good. Between these two lies a gap we currently lack the ability to fully bridge.

IV. Economics: What Happens to the Market When Costs Approach Zero

Let's jump from micro-mechanisms to macro-structure.

In economics, when a resource can be used for free, the rational choice of each individual leads to collective disaster.

When marginal costs approach zero, what happens to the supply curve?

Traditional content production has barriers: you need time, experience, and expertise. A valuable technical article might take an engineer with ten years of experience three days to write. This cost forms a natural quality threshold—things not worth writing, no one will spend time writing.

LLMs have compressed this cost to nearly zero. Now, anyone can produce a 5,000-word "in-depth analysis" in ten minutes.

The result is predictable: the supply of those capable of producing high-quality content has barely increased, but the supply of those motivated to produce low-quality content has exploded.

The AI/computer science field is the epicenter of this disaster, and it gives me a headache. This field once had barriers—you needed to genuinely understand the technology to write something valuable. This threshold filtered out a large number of people just chasing traffic, since they couldn't write.

Then AI heated up and became a traffic goldmine. The emergence of LLMs made the threshold disappear.

All sorts of people came sniffing around—marketers, operations folks, self-media influencers—frantically churning out "10 AI tools that will 10x your productivity." News outlets started producing "ChatGPT will replace all programmers within three years." Every day, opening Douyin, Zhihu, or WeChat official accounts, self-media personalities began reciting every sentence from Karpathy's podcast in an endless barrage, but stripped of all detail and reasoning, leaving only conclusions.

Where are the real practitioners? They're still there. But the environment has gone bad—writing genuine articles doesn't pay off like chasing traffic does. This really frustrates me, because when I search for a technical problem, the top 10 results are all posts from big influencers selling "skills" to earn traffic, with zero originality. What I need is a real technical pitfall someone actually encountered, not your daily "xxx skill" spam.

This is how the drowning mechanism works: not by eliminating good content, but by overwhelming it with quantity.

V. We're Already on This Road, and There's No Turning Back

I need to be honest: this is not an article that offers solutions.

Because I believe that, given the current technological and social structures, there is no systemic solution to this problem.

Search has essentially become ineffective.

LLM-generated content has an advantage that hasn't been solved yet: it's SEO-friendly. It naturally distributes keywords in the right places, has clear heading structures, and hits the right word counts. Search engine ranking algorithms will, for a considerable time, continue to rank this type of content at the top.

I have no conclusion.

I just want to record a feeling: when we search the internet for genuine experience on a specific problem, it's becoming increasingly difficult. Not because there's less information, but because there's more information, and the real signal within it is being diluted.

This isn't a technical problem. It's an ancient problem of human society, just wearing a new shape.

When something becomes cheap, it gets abused. When it gets abused, it depreciates.

Content is going through this process right now.