Even the heaviest action goes no further than “reducing visibility”; we avoid outright deletion wherever we can—because every step further an action goes, the harder the cost is to undo if we get it wrong.

Seven degrees of force, one irreversible spectrum

They’re all called "moderation," yet how reversible they are differs enormously. Left to right, the edge shifts from green to red, and the cost climbs from "still recoverable" all the way to "irreversible." Hover any card to see where it lands on the spectrum.

The quiet room: the key invention that worked out by accident

The quiet room (April 2023) came into being almost against its designers’ intentions—the product of two separate needs that happened to converge. First, operations had to pull articles out of the feed channel by channel, by hand: “exhausting; we need a one-click fix.” Second, faced with troll-army flooding, the thinking was, “rather than hide posts one at a time, just block the whole account—and the new accounts they spin up pad our sign-up numbers anyway.” With the quiet room, most flooding could simply be “kicked in there,” which made moderation lighter, and easier to reverse.

An unresolved tension: account-centrism vs. work-centrism

Operations asks, “has this account broken the rules?” Product asks, “is this piece any good, and might this person change?” For a long time the system has deliberately refused to define “who is a spammer,” working instead by time windows, re-evaluating afresh each period, never pinning a permanent label on anyone. The catch is that while “people can change,” there has never been a matching mechanism to “let them back in.” This is a mark left by a real organization inside the messiness of governance, and it is unsolved to this day.

One platform, three kinds of content, three cleanup regimes

Even when it is all spam, long-form articles, moments (short posts), and comments do not harm users in the same way, so the platform’s moderation flows split accordingly. The judgment rests mainly on three things: how much signal there is, how fast and wide it spreads, and how “invasive” it is toward others. The treatment fits the ailment—one set of rules is never forced onto everything.

Long-form takes the “noise-reduction” path plus the quiet room. From the very start, article governance has relied on machine detection to adjust visibility automatically, on the principle “don’t delete the content, just stop it being seen”: a suspicious article with a high score no longer appears in recommendations, trending, tags, or channels, but its content, its page, and its existing links are all kept. For a censorship-resistant platform built to protect Chinese-language writing, this holds the cost of “suppressing spam” at “reducing reach” rather than letting it become “destroying content,” and it leaves a recoverable way back for political or academic articles the model wrongly flags. These actions are mostly automatic and silent, operating on a single article.

Moments lean on community design, and less on the model. Short posts carry little text, plenty of ambiguity, and hard-to-grasp context, so judging by content alone easily goes wrong—false positives are especially troublesome. In the past, moments each sat on their author’s own page rather than being pooled together, so even an offending account blasting out ads would not flood anyone else’s feed; but the moment you gather them into a single channel and rank them by engagement, mindlessly poured-in spam would “die on arrival.” So the “Chit-chat” channel switched to an opt-in Closed by default; anyone who wants to take part has to actively apply and pass review before they can join. + buffered admission: to join you first tap “Join,” and you are approved automatically once a moderator has looked it over and found no problem, or if no one rejects you within 48 hours. Within half a month of opening, “Chit-chat” had cut spam to zero—at a time when spam short posts made up more than half of the whole site. Beyond the model, it relies even more on catching “rings” (a single operator mass-producing accounts and flooding with one shared template) to sweep them all up at once.

Comments are the most invasive, and the actions against them the most aggressive. When everyone posts their own short pieces, the streams stay out of each other’s way; but a spam account will “actively” turn up in the comments beneath someone else’s ordinary article—the user cannot avoid seeing it, and the notifications keep pinging. This is the single most complained-about experience. So comments stack four layers of defense: volunteer cleanup, automatic model cleanup, spam-ring (Ring) detection, and moderator judgment. To avoid sweeping up legitimate comments by mistake, automatic cleanup sets very strict conditions: the same content must recur three times and match a ring pattern, or the model’s score must be high enough to be near-certain; otherwise it holds off and hands the call to a member of the neighborhood-watch team. The action operates on a single comment, and usually removes it on the spot (unlike articles, which are merely down-ranked).

Further reading: Who counts as a spammer? Account-centrism vs. work-centrism