The Reading App Where Everyone Now Suspects the Reviews Are AI

WeChat Reading is the biggest book app in China, and its most distinctive feature sits beside the books. It is the layer of other people's highlights and comments that runs alongside the text, so you read a novel with a few thousand strangers reading it over your shoulder.
For years that layer was the good part. This month it became the argument.
The trigger was small. A reader recommended an Agatha Christie novel to a friend, and the friend reported back that the app had shown the ending before the first chapter, because the comment layer surfaced discussion of the plot on the way in. That complaint, posted on 11 November, was widely shared, and it is the least interesting of the complaints circulating.
The accusation that runs both directions
The day before, a different post made the opposite case: the comment layer had filled up with reviews that read like they were written by a model. Formulaic openings, tidy three-part structure, an emotional turn in the last paragraph that resolves nothing. And then the part that stings: readers who wrote carefully were being accused of being AI.
That is the new social dynamic. Once enough text in a space is generated, the burden of proof moves onto the humans. Style becomes evidence. A well-organised paragraph is suspicious. A neat summary is a tell.
In a space built on the assumption that a comment is a person thinking out loud, this is corrosive in a specific way. The comment layer's value came from the fact that someone bothered. If that can be faked at scale, the only remaining signal is how much the writing costs the writer, and nobody can see cost from a screenshot.

The metrics made it worse
WeChat Reading does more than show comments. It computes a recommendation score, and it maintains a leaderboard of highly rated books. A review with a high score can push a book onto that list. In August, a 973-character review was reported to have reached a recommendation value above 92%, which put the book in front of a lot more readers.
A metric like that creates an incentive that has nothing to do with reading. If a well-structured, emotionally satisfying, slightly vague review scores well, then producing well-structured, emotionally satisfying, slightly vague reviews is a rational strategy, whether you are a person or a script. The output of that strategy is indistinguishable from what readers are now calling AI flavour.
The app also shipped an AI feature in May, which the company framed as an assistant for readers. The existence of that feature is the problem's context rather than its cause: it gives every accusation a plausible mechanism. You do not have to prove that a review was generated; you only have to note that it could have been.
That asymmetry is what makes the argument unresolvable. A reader who suspects a review can point at the app's own assistant. A reader accused of being a machine has nothing to point at, because there is no proof of authorship in a text box.
Then there is the author
In October, a reader gave a book a middling rating and received a direct message from the author asking for it to be changed. The screenshot drew more than 1,700 likes and several hundred comments, and the reaction was mostly about the asking, not the rating.
Put the three stories together and a pattern emerges. The comment layer was designed as a social space, was measured like a product metric, and is now being policed like a public forum. Each of those framings implies a different fix, and the app has been choosing the third while running the second.
What the readers are actually asking for
Read the complaints closely and the demand is narrower than "remove AI". It is roughly: do not show me other people's reactions before I have had my own, and if you rank those reactions, tell me what the ranking is for.
The spoiler problem has a clean fix, which is a reading-progress filter that hides comments beyond the page you are on. Some apps do this. It costs a toggle and a data model, and the data model is the hard part: the app has to know where each reader is in each book, and then decide what a comment anchored to page 240 is allowed to reveal to someone on page 30.
WeChat Reading already tracks that position, because it syncs progress across devices. The information is there. What is missing is a rule about who is allowed to see what, and rules like that are political inside a product team, because they reduce engagement on exactly the screens that engagement is measured on.
The AI problem does not have a clean fix. Detection is unreliable, accusation is cheap, and the cost of a false positive falls entirely on the person accused. A platform that starts labelling reviews as generated will spend the next year arguing about the labels.
The more interesting option is to stop ranking reviews at all. Recommendation scores turn a conversation into a leaderboard, and leaderboards reward exactly the qualities that make text look generated. A space that only sorted by recency, or by how many people highlighted the same sentence, would be worse at surfacing the best review and much harder to game.
The uncomfortable version
There is a version of this story in which nothing was generated and the whole panic is about a style. Chinese book-review writing has its own conventions, and those conventions happen to overlap with what people now recognise as model prose: the symmetrical structure, the tidy conclusion, the emotional beat placed at the end.
If that is what happened, the lesson is worse than the AI one. It would mean the house style of online review writing was already machine-shaped, and the models learned it from us. Readers are detecting a genre, and that genre has been getting more mechanical for a decade.
Either way, the comment layer is now a place where readers watch each other for tells. That is a strange thing to have built beside a library. The books did not change. The margin did.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.