When AI-generated texts are flooding everywhere, is there still hope for quality content?

These days, the hashtag #Tangjiasanshao says AI content spinning has become a malignant tumor in the industry# has topped the trending search list. The cause is his signed article "Let AI Content 'Show Its True Identity'" published in *Guangming Daily*.

He stated in the article: "In the past, plagiarism was 'copy-paste', which could be spotted at a glance. Now with the help of AI, plagiarism has turned into 'deconstruction and reorganization' — taking the framework of hit works apart, using AI to fill in the details again, and generating a 'brand-new old article'."

Image source from Weibo

How fast can AI be?

"It can generate 5 million words in 48 hours. The scale is so large that it gets out of control: the number of new debut books on one platform has soared from 400 to 5606 within a year. When plagiarism can be produced industrially, original content becomes a luxury. Who would still be willing to spend ten years polishing a single work?"

To this end, he further called for "filling the gap of AI identification in the text field, so that AI content can 'show its true identity'."

This article has a good original intention, and the appeal is also very pertinent, but it also makes "TopKlout" fall into deep thinking. Can completing the AI identification of text content really solve the problem?

01

Online Literature Is Just The First Collapsing Starting Point

The reason why Tangjiasanshao complains about the high AI content in online literature is not only that he is an online literature author himself, but more importantly, the business model of online literature relies too much on "word count".

In this industry, updating 10,000 words per day is a hard requirement. In the past, relying on human authors, not to mention how well the plot is written, just typing 10,000 words would take a lot of time. But after the emergence of AI, this matter has become extremely easy. Because what AI is best at is exactly this kind of "large-volume, regular, formulaic" content.

Just like the example Tangjiasanshao cited in his article — as Mao Zhihui, vice chairman of Jiangxi Online Writers Association, said — his limit is updating nearly 10,000 words per day, but in front of AI that can generate tens of thousands of words in a few minutes, he is directly hit by "dimensionality reduction attack".

A human being cannot compete with AI in production capacity by nature.

Moreover, the proliferation of AI is no longer new, it is just named by Tangjiasanshao now.

Those "hit articles" on official accounts with perfect structure, full of golden quotes, but leaving you nothing to remember after reading; those formulaic, routine product recommendation notes on Xiaohongshu that share almost the same first three sentences; even a lot of purely human-written text now is so influenced by AI that it is hard to tell the real from the fake.

His statement that "online literature is the hardest-hit area" is still one-sided. From the perspective of "TopKlout": All places where text can be distributed are hard-hit areas.

Online literature is just the first field that collapses, because it is closest to money, has thinner barriers, and most users care more about the pleasant reading experience than writing skills and logic. The invasion of media articles and classic literature by AI is also happening right now.

To be honest, many media outlets that the author likes, whether official or self-media, almost 99% of their content has traces of AI.

In early June this year, Hugo Award winner Hao Jingfang even admitted frankly that in her new book *Galactic Academy*, "AI writing has accounted for half of the total content, and readers cannot tell which parts are written by AI".

Compared with most online literature, these two types of content are more carefully produced by more professional people, but they have already had AI intervention, and the general trend is irreversible.

Tangjiasanshao said that "AI content spinning is essentially a systematic strangulation of the spirit of innovation". To put it more bluntly: when plagiarism is more efficient than original creation, what gets eliminated is not plagiarism, but the motivation to create original content.

A writer finds that spending ten years polishing a work cannot get the same return as updating 30,000 words per day, he may stop writing forever. When people who want to write no longer create, there will be more and more garbage produced by AI.

02

Why the "Labeling" Route Does Not Work

Now that AI is so widespread, does it make sense to let AI content "show its true identity"? It makes sense, but the significance is limited.

First of all, technically, text is naturally not suitable for watermarking.

Images can embed metadata, videos can write marks into codes, which users can easily notice, and unmarked content is easy to be reported, but text cannot do that.

Tangjiasanshao also mentioned this point in his article: "Text identification marks are easy to be deleted or tampered with in batches. A large number of AI-generated text content is spreading 'invisible', which cannot be perceived by users or traced back by platforms. Therefore, the top priority is to learn from the previous experience in the video field, and actively explore ways to identify AI-generated text content."

Active exploration is certainly good, but no one knows when we can get the final solution.

At present, the industry consensus is: The accuracy of general text detectors is not sufficient as a basis for judgment. Because the principle of many detectors is to count language features — sentence regularity, word distribution, punctuation habits.

For example, this article written by Tangjiasanshao has a large number of double quotes, dashes, smooth logic, regular paragraphs, and many words similar to AI-generated expressions... If we take the detector result as the standard, it is very likely to be judged as written by AI.

Furthermore, there are countermeasures against policies.

Many AI-generated articles will add prompts such as "use more colloquial expressions, use more short sentences, and occasionally repeat" to avoid looking like AI-generated content.

Many AI tools have the "de-AI-style" skill. Many authors will even deliberately write typos in the published articles to make the content look manually labeled.

Screenshot source from WorkBuddy

Second, in terms of mechanism, labeling is an adverse selection.

Letting creators voluntarily declare "I used AI in this work" is logically equal to letting a thief write "I am a thief" on his forehead. Would people who want to use AI to get traffic in batches be willing to label their content voluntarily?

In April this year, the Xiaohongshu AI Governance Open Day disclosed that since the test of governance rules, more than 120,000 pieces of content have completed voluntary labeling. 120,000 pieces sounds like a lot, but in contrast, Xiaohongshu has disposed of more than 1 million cases of AI improper behavior since 2026, including 800,000 AI hosting accounts.

The gap between the two sets of data is huge.

Therefore, the creator's own declaration does not have much effect, which requires the platform to have corresponding judgment capabilities. But this goes back to the original dead cycle: it is too difficult to judge whether a text is written by AI at the present stage.

Third, it is difficult to measure "how much AI usage counts as AI-generated content".

Is a text that uses AI to polish one sentence counted as AI-generated? Is a text that uses AI to build an outline and is finished by human counted as AI-generated? Is a text that is dictated by human and sorted out by AI counted as AI-generated? Is a text that uses AI to check typos and modify wrong sentences counted as AI-generated?……

The boundary between AI and human labor is very difficult to define clearly. There have indeed been real cases where a text that only uses AI to check a few typos and modify two sentences was judged as an AI-generated article.

Of course, major AI manufacturers have also thought of solutions. In August this year, Anthropic officially launched a statistical text steganographic watermark, and all outputs of the new version of Claude are marked by default, which cannot be turned off.

Steganographic watermark means that a single sentence cannot show any abnormality, and human reading is completely unaffected; but when the text reaches a sufficient length, such as hundreds of words or more, this continuous weak probability offset will form a statistical fingerprint that can be verified by a key, that is, an invisible watermark.

This event triggered huge discussions as soon as it came out. To a certain extent, it can indeed tell the machine which content is AI-generated.

But many creators believe: I just use the tool for polishing, which does not mean that I agree to embed a permanently traceable mark that can be verified by third parties in my text.

Some journalists, academic authors, and people who write sensitive content will feel that the text will carry traceable meta signals, leaving text fingerprints and bringing traceability risks.

03

AI Is Not The Problem, Value Is

At the end of the writing, we might as well think from another angle: what we hate is AI, or low-quality content?

You click on an article, read 800 words, and get nothing; you scroll to a video, watch 30 seconds, and find it is a spliced fake news; you click on three pieces of content that "look like dry goods but are full of nonsense" in a row, and you will not click the fourth one. What is consumed is not only time, but also credit limit.

Analyzing from this perspective, what many people care about is not that AI takes away traffic, but that garbage content takes up attention. When people read too many AI articles and AI pictures, it won't be long before they can't remember what purely manually created content looks like.

At this point in time, we still have the ability to protest, but after a year or two, when we get used to AI sentence patterns and our aesthetics change, how many people will care about how this content is generated?

Therefore, platforms might as well change the question from "whether to label" to "whether to recommend".

A report completed with AI assistance but containing on-site interviews by journalists and exclusive data should be recommended as long as the content is good; while an article typed manually by humans but full of meaningless platitudes should also be limited in traffic.

Replacing the identity standard with the quality standard may be a more feasible route. But another problem is that the recommendation logic of many platforms has long been incomprehensible to people. (Refer to the previous article of TopKlout Official Account: Do we need such a powerful "algorithm recommendation"?)

Nowadays, it is impossible to completely abandon AI. How to coexist with AI and tap the valuable value of AI is the meaningful thing to do.

Why do I say that?

Because the post-2000s and post-2010s generations are the last group of people who know what the world was like before the emergence of AI. The post-2020s generation, who were born only a few years ago, will be exposed to almost all text that is "polluted" by AI in the future.

What is the point of blindly distinguishing and blindly resisting?

Several times when I was listening to songs and reading books, I would be in a trance and full of emotion: Will my children be able to feel the real emotions in the text written by real people in the future?

04

Conclusion

Is labeling meaningful? Yes. It at least establishes a direction: AI-generated content should not disguise as human expression.

But can it solve the problem? No. At most, it gives readers a reminder before they click, and cannot change the content production process.

In summary, what can save high-quality content is never labeling AI, but letting valuable content get more recommendations.

This article is from WeChat Official Account "TopKlout" (ID: TopKlout), author: Xiaosong, published with authorization from 36Kr.