The Hollowing of Content Industry: When AI Produces a Pile of Air

This article is a bit unusual. Not because something earth-shattering happened, but quite the opposite — because nothing happened. I was supposed to have a draft of upstream output, and the entire pipeline ran to completion. The status code was normal, none of the stages reported an error — but the single most important part, the content itself, was a blank.

The review JSON for the critique-A step writes it more directly: in missing_constraints it says "Upstream 'drafting A' produced no draft content at all, so critical review cannot be performed; there are no falsifiable claims to analyze."

I'm reading that sentence verbatim because it's so precise. The whole pipeline is running, but it has not produced anything of value. Every step executed, every status code normal, but the final output is zero. This isn't a bug; this is structural idling.

We are mass-producing content that looks complete but is actually hollow

I've seen too many AI content teams, and they fall into roughly three categories. The first is one that actually gets value from AI — clear input standards, strict output validation, and well-defined human-machine collaboration boundaries in the workflow. The second is one that uses AI noisily but produces poor output — a long list of tools, beautifully built knowledge bases, but open any output and the quality is below the passing line. The third is what I recently encountered — the pipeline is running, the people are working, time is passing, but in the end no actual content gets produced.

The third type is the most extreme, and also the most easily overlooked. Because its appearance is the cleanest—no obvious errors, no user complaints—only that final "empty" state, like a small thorn stuck deep in the process. You don't feel it most of the time, but it's the wound that's actually bleeding.

Why do I say this problem is more serious than "AI quality isn't good enough"? Because it exposes a deeper mistake—we equate "using AI" with "producing content." This logical fallacy sounds low-level, but I've seen too many teams fall on it. They have AI knowledge bases, AI research reports, AI-generated daily summaries—looking like peak productivity, but when you actually open the content, it's either universally correct platitudes, or—as in my case—nothing at all.

There's one thing I have to say straight out here: an AI spinning its wheels is more dangerous than no AI at all.

Without AI, a team would clearly know that it had no content, and would manually supplement it, or directly admit that this area was beyond its ability. With fake AI, the team thinks it's producing content—archiving on schedule, sending on schedule, meeting on schedule—but in reality nothing is being produced. Everyone is busy, everyone is consuming time, no one feels there's a problem, until one day looking back—this half year, we've produced a pile of air.

Cross-domain collision: why this problem is exploding right now

Why is this happening at this point in time? I think there are three structural reasons that happened to collide today.

The first is the explosion of tool richness combined with the absence of verification mechanisms. Before 2023, AI content tools were still relatively primitive, and when people used them, there was at least an implicit manual check — because the tools weren't easy to use, humans would look at the output an extra time. But by 2025, the toolchain had become extremely rich, with ready-made AI solutions for every step, from knowledge bases to research to podcast scripts. This richness actually eliminated human attention — because every step had AI, every step assumed the upstream was reliable, and so the whole chain became an automated machine where no one truly looked at the output.

The second is that speed worship has replaced quality worship. I've met many AI content teams, and they evaluate workflows on essentially one metric: speed. Content output is three times faster than before, so the toolset is "good." But speed is an input metric, not an output metric. A machine that produces a thousand screws a second—yet nine hundred and ninety of those screws are defective—has a speed of zero.

The third is something I've observed recently: the AI content industry is going through a "standards vacuum." Traditional content industries have editorial boards, peer review, and audience feedback — these mechanisms naturally weed out low-quality content. But the AI content industry still measures output by "production speed" and "format completeness" — the two dimensions at which AI is most adept at faking it. You can make hollow content look perfectly formatted, making it appear more like "high-quality output" than content that has a point of view but a slightly rough structure.

These three reasons collide at this point in time, in mid-2025, generating a massive amount of systemic idling. Here's a specific bet I'm making: within the next 18 months, we will see the first concentrated "black-box reckoning" in the AI content production field—not one company collapsing, but a batch of teams simultaneously realizing that their AI workflows have actually been idling for the past six months. The trigger for this reckoning, I suspect, will be a cliff-like drop in the user retention rate of some leading AI content tool, because users finally discover: after half a year of using AI, content output has gone up, but quality has gone down.

Naming Ceremony: What We Call This Kind of Failure

At this point, I want to do a small experiment at the language level—because I think this problem is serious enough to need a precise name, rather than being brushed off with vague descriptions like "AI didn't do a good job."

I considered a few candidate words. The first is "process inflation"—just like monetary inflation, more and more processes are in circulation, but each process buys less and less. The strength of this word is that it captures the essence of the problem: it's not that there aren't enough tools, but that the "purchasing power" of the tools is dropping.

The second candidate word is "false-busy trap"—sounds more colloquial, easier for practitioners to accept. The state it describes is very accurate: a team looks busy, every part of the operation is running, but in fact there is no meaningful output.

The third candidate word is my favorite: "the hollowing-out of the content industry". The word "hollowing-out" precisely describes what's happening—the structure is still there, the shape is still there, the volume is even growing, but the core is empty. Like a hollow tree, looking lush from outside, it falls with one gust of wind.

I chose "The Hollowing-Out of the Content Industry" as the main title of this episode because it frames the problem at the industry level, not as some specific operational mistake. This is not a tool bug; it's a systemic industry-wide bias. We are using AI to mass-produce a new kind of waste—not material waste, but attention waste. This content doesn't hurt anyone, but it consumes readers' time, consumes editors' attention, and consumes the entire industry's trust in AI content tools.

Self-dialogue: If Someone Tells Me I'm Overthinking

I know some will say: Bian Yang, aren't you overreacting? One upstream failure, and you wrote a whole article about an industry crisis — is that being a bit paranoid?

I think this rebuttal is valid. I have to admit that the core evidence of my article is a single upstream idling event. Of course, this could be an isolated bug—the upstream AI tool happened to encounter an edge case it couldn't handle that day, the output was zero, and the downstream critique had nothing to critique. That's a normal technical glitch, not representative of the whole industry.

But, and I must say "but": it is precisely this mindset of "explaining everything with technical glitches" that is itself part of the problem. If every AI hallucination is attributed to "the tool's occasional issues," we will never have the motivation to build systematic quality assurance mechanisms. Like food safety—if every case of food poisoning is explained as "this batch of ingredients happened to be problematic," regulators will never have the motivation to establish routine inspection systems.

Others may say: the hollowness of the AI content industry is just a passing pain. Today's AI quality is not good enough because the technology is still in its early days; in another two or three years, models will be more powerful, the content they produce will be of higher quality, and the hollowness problem will naturally disappear.

I partly agree with this argument. It is true that AI's capabilities have improved, and the pace of model progress from 2023 to 2025 has exceeded most people's expectations. But that is precisely what worries me more — if even today's model capabilities are not enough to make our content workflows produce effective content, then the problem is not in the model, but in the design logic of the workflow itself. A workflow without an inspection mechanism, no matter how strong the model becomes, is simply producing more polished air more efficiently.

So I stick to my main thesis: hollowing-out is not a technical problem, it's a mechanism problem. What we need now is not stronger models, but stronger input standards and output verification.

Personally Verified: A Simple Way to Test Whether AI Content Is Really Producing Value

Speaking of verification, I want to give you a method I've personally tested to combat the hollowing-out of the content industry.

In my content production workflow, I now require that every AI production step, before "claiming completion", must satisfy a specific, verifiable standard — not "content has been generated," but "this content must be able to answer at least one specific question whose answer you didn't know before."

This standard looks simple, but it acts like a filter that instantly exposes every piece of AI output that "looks complete but is actually hollow." I tested it on myself: the moment I added this rule to an AI research report, I immediately spotted the problem — a lot of what I thought was "producing content" was actually just repeating things I already knew, or repackaging vague concepts in different words. But with this checkpoint in place, the AI is forced to be more precise — it has to give me concrete data, concrete sources, concrete path judgments, instead of generic "worth watching" and "keep an eye on it."

This is essentially a replica of an old engineering principle—Test-Driven Development (TDD)—applied to the content production domain. The only difference is that in content, our "test cases" are not code coverage, but a question that can be concretely answered: For the domain this content claims to cover, what previously-unknown-to-me questions can it answer?

If an AI-produced content can't answer a single specific question within the field it claims to cover, then it's empty—no matter how beautiful its format, how long its length, or how cutting-edge the topics it covers.

Concrete Actions You Can Take Today

Having said all that, I want to leave you with a concrete, actionable move at the end, rather than vague calls to "stay tuned" and "keep watching."

Do not chase the next new AI tool. Do not compare which model just updated. Do not read the next "AI-era content production guide." Open up your team's AI-assisted output from the past week right now, pick any three at random, and see how many concrete questions — ones you didn't already know the answer to — each one can answer.

If two out of three of those prompts go unanswered, you don't need a different tool. What you need is to go back to fundamentals — tighten your input standards, build out your verification mechanism, and let AI truly take on amplifying tasks in your workflow, not substituting ones.

Specifically, you can establish a simple validation rule starting today: Every piece of AI-produced content must, before being judged "complete", be read by a human who actually annotates at least three specific valuable knowledge points. Content with no knowledge points is not archived, not sent, and does not count as output.

I know this sounds dumb, and it slows things down. But it is precisely the most effective medicine for treating hollowness — not a stronger model, not a faster tool, but the simplest possible mechanism: making every step of output pass through a human reviewer's check that asks, "Is this useful?"

This question is worth thinking about carefully for everyone who is using AI to get things done.