Labeling AI text may be a trap writers are setting for themselves

On September 9, the novelist Tang Jia San Shao (Zhang Wei) published an essay in Guangming Daily.

His complaint is concrete. AI content laundering is no longer the copy-paste of years past. The current method is to strip a hit novel down to its skeleton, have an AI pack new flesh onto it, and ship a brand-new old book. Five million characters in forty-eight hours. He has a name for the output: digital slop.

He makes three proposals — stronger industry self-regulation, better legal standards, and a focus on protecting minors. The most urgent, he argues, is to close the gap in labeling for text, so that AI-written work declares itself.

The demand is easy to sympathize with. China’s Measures for Labeling AI-Generated Synthetic Content took effect last September, and on the video and image side Douyin, Bilibili and Kuaishou have all shipped AI labels. Text has nothing. Why should it be exempt?

I want to make two points. Both run against the grain of most people’s first reaction.

1. The premise has a shelf life

Demanding a label on AI-generated text rests on an assumption nobody states out loud: that readers who see “AI-generated” will steer clear.

Put differently, human writing is more competitive, so marking the difference helps humans.

Today that assumption probably holds. The problem is that it expires.

Labeling regimes have one defining property: once established, they do not come down. A rule does not lapse because the technology improved. You demand the label today; five years from now, when the machines write well enough, you still have to carry it, and by then its meaning has flipped.

On that day the label stops being a shield and becomes a comparison table. A reader scrolls past two pieces, one labeled AI and one not, and discovers that the labeled one is just as good, maybe easier to read. Multiply that comparison by a few million a day and the outcome is not hard to picture.

What writers are fighting for right now is a credential for their future competitor, issued while they still hold the lead.

2. Even a perfect label never touches the real problem

This is the deeper layer.

Five million characters in forty-eight hours. Sit with that number. It tells you how far supply has expanded.

Now look at the other end. How much can one person read in a day? That figure has not moved in years. Attention, patience, and comprehension did not scale with the presses.

So the scarce thing was never knowing who wrote a passage. It is knowing whether the passage is worth finishing.

A label serves the right to know: am I being deceived? It does nothing for attention: do I have the time? Even if every one of those five million characters carried an honest AI tag, all five million still clog the feed and still push the good work further down.

Treating labels as the cure means answering a flood on the supply side with a tag on the receiving side.

3. So what would actually work

Having rejected two things, I owe you something constructive.

First, labeling and anti-laundering are two different problems, and bundling them means neither gets solved.

Labeling governs the right to know. Laundering damages copyright. Even if every AI text declared itself, not one instance of skeleton-and-flesh would disappear. The harm in laundering was never that readers mistook it for human work. The harm is that it routes around the original author’s rights.

Of Tang’s own three proposals, the second is the one aimed at laundering: define the threshold for original expression and settle who owns what. That one is far harder than labeling, and far more important.

Second, if we must label something, we are labeling the wrong thing.

A more informative question than “did an AI write this passage” is “is anyone accountable for this judgment.”

Someone drafts with AI, then verifies it, signs it, and eats the consequences when it goes wrong. I trust that. Someone else types every character by hand, but the content is carted in from elsewhere with no judgment of their own anywhere in it. I do not trust that.

The method of production is a weak signal. The party who answers for the result is a strong one.

Last

Tang is not asking the wrong question. The phenomenon he describes is real, the anxiety is real, and five million characters in forty-eight hours should put a chill down the spine of anyone who writes for a living.

I only think that labeling, as an answer, cannot deliver the protection he wants in the short run and may bite back in the long run.

What has always protected people who write is not a tag. It is a judgment nobody else could have produced.