Further reading · Platform governance

Who gets judged to be a spam account (spammer)

Most platforms define the spam account (spammer) first, and then handle accounts according to that definition. For years Matters never wrote this definition down, and the reason is not simple oversight. Starting from a single internal discussion, this article lays out how the platform makes its calls across accounts, works, time slices, and manual handling — and which problems this approach leaves unresolved.

A word that was never defined

Almost every content platform treats being a spammer An account — or the person behind it — that mass-posts spam such as ads, scams, and pornography. as a computable state. Posting frequency, similarity, the number of outbound links, the number of times reported — cross a certain line and the system stamps an identity onto the account. It is the natural thing to do in engineering terms, and it is exactly the thing Matters has never done.

“This is also why we've never marked spam out in the open — nowhere in the system will you see us defining a person as a spammer.”

The back end has a spam-probability score, a blacklist, and even a one-click “quiet room” that pulls an account out of the public space — everywhere you look, it seems to be sorting accounts. Yet behind all these actions, no numeric threshold for “how far does behavior have to go before it counts as a spammer” has ever been applied directly to deletion or to exclusion from the homepage ranking. Deletion and the quiet room are both case-by-case judgments made by operations, not conclusions the system computes on its own.

Push a little further and this “refusal to define” traces back to the philosophy of an early product manager. To avoid naming a real person in an internal discussion where identities are already known, we refer to that role below as “product,” and keep the pseudonym Benson used in the discussion. Product's position is paraphrased as follows.

“The posts this author puts out may well be spam, but the person may change, so we've never set a threshold for exactly what behavior makes someone a spammer — that's the thing we've never defined.”

“People can change” is a premise that was never written into the code, yet has long shaped the logic of how cases are handled. It keeps the system from permanently defining an account with a single label, and it also loads more manual judgment onto operations in their day-to-day work. What comes next is how this premise gave rise to two different approaches, and the limits it runs into in practice.

Time slices: spam is a state of the current period, not an identity

If you don't stamp a permanent label onto an account, how does the back end ever compute those figures for “ spam Worthless content — ads, scams, pornography, content-farm material — mass-posted to make money, deceive people, or juice traffic. account count” and “spam article count”? The key lies in the unit of counting. These numbers are not a permanent list; they are computed as time slices.

“It's month by month that we look at the quality of their posting to decide about this person … it's not that I've already tagged this person as spam and then excluded them, no.”

In other words, the judgment of “spam or not” is pinned onto “this account, in this month / this week,” not onto the account itself. The precondition for a judgment is that there must first be posting behavior — as it was put, “they say you made a move, and only from your move do we judge.” An account that goes on a spam spree this month is spam in this month's slice; if it then falls silent for the next half-year, it counts as neither spam nor a pollutant in any of those months' slices, because it simply made no move to be judged on.

Someone in the discussion pressed further on “whether the same account gets counted more than once.” The answer is yes. The system only judges behavior within a given time interval, so an account that posts junk every month shows up once in each month's data. This kind of counting is closer to a series of snapshots than to a permanent list.

This yields a result that looks counterintuitive but follows the rules. An account judged highly spammy half a year ago can, as long as it makes no move this month, have a score of “zero” in the latest month's data, and be treated as a normal account. Someone in the discussion confirmed this plainly: the rules “will be rewritten,” because the unit of judgment has always been current-period behavior, not past identity. So the premise “people change” shows up in the data as a re-evaluation every period, rather than as a permanent mark.

The crack between belief and implementation

The problem is that if “people change” is a design premise, there ought to be a matching provision to go with it. When an account's score turns good again, in theory someone should let it back out of the quiet room. But the most pointed exchange in the discussion reveals that this, in practice, has never happened.

Operations' working logic is direct and pragmatic: the quiet room is permanent — “once you're put in, you don't come back out.” So one member laid the contradiction bare.

“If they really do score high later, shouldn't we let them out? Otherwise why bother scoring them this way at all?”

The answer was that the later scoring “isn't actually used for anything.” Product believes people change, and so refuses to hand down a permanent verdict; operations has to block the harassment in front of them right now, and so, once someone is put in, never looks back. One side preserves the possibility of “turning over a new leaf,” while the other never prepares any follow-up action for that possibility. The result is that the two logics each hold on their own terms yet hollow each other out: the score is recomputed every month, in theory giving people a chance to change; the quiet room, once shut, never reopens, so in practice no one is ever actually let out. Since there is “no follow-up action” telling anyone to release people, whether or not to score them, and whether or not people get better, makes no real difference to the daily work.

The point of preserving these discussions is to show the day-to-day of a real team — and, with governance experience accumulated, to see anew the places where an early belief, an engineering design, and daily operations never quite joined up.

Account-centrism vs. work-centrism

To make sense of this difference, the people in the discussion reached for a pair of contrasts. This pairing captures the two ways of judging that have long coexisted in Matters' handling of spam.

Both positions have their own rationale; they just look at the same thing from different places in the system. Operations deals daily with “blocking harassment”; product thinks about “how to make the homepage better while not stripping anyone of the right to speak.” The people in the discussion ultimately traced the divergence to the organization's own history: at one point it was more heavily staffed, with a fine division of labor, and the same problem, absent enough communication, had two separate lines each grow its own solution. This looks more like a governance mark accumulated as an organization grew than like a single functional bug.

But if you look only at the substantive effect, the two approaches end up very close.

“The quiet room = the algorithm tags it as a statement and keeps it out of the ranking; substantively the same as suspension, the only difference being whether the account gets a quiet-room label.”

That is, work-centrism's “down-weighted to invisibility” and account-centrism's “suspension” look almost identical from the user's side. The two approaches differ in their reasons, but their external effect is close: this person has vanished from the public space. The difference mostly stays in the system's records — whether or not a note gets kept that “this one is a spammer.”

The boundary of “obviously junk”

Account-centrism wins out in daily practice because most cases look “obvious.” But the people in the discussion also admit, again and again, that this “obvious” is extremely fragile at the edges. The most honest confession comes from a case where someone was put in the quiet room who was in fact a real person.

“I really couldn't tell whether it was content or it was an ad … if they turned around and challenged us — saying, I clearly wasn't posting ads (even though the text itself repeats a lot), what gives you the right to put me in the quiet room — then honestly the platform might not be able to stand fully, one hundred percent, on solid ground.”

Gray zones like this recur throughout the material: the “moving-house post” that wraps everyday jottings up front with the ad hidden at the back; the “financial news share” that smuggles in the author's own brokerage; and a kind of AI flood-posting that has only appeared in the last couple of years — the content doesn't read badly, and the author can put out five, six, seven, eight pieces a day, flooding the whole board with their own stuff. For this last kind, the people in the discussion admit that their eventual reason for acting is usually “you're flooding way too outrageously,” not “you broke a rule.”

“In the end the reason they get put in is always that they're flooding … the boundary here is really blurry.”

The material also shows a practice closer to a stopgap buffer: the team once herded a large volume of content that “breaks no rule but doesn't belong on the main public pages” into a specific topic channel. When the judgment of “obviously junk” in fact mixes in the poster's vibe, their posting frequency, and their writing style, that seemingly clear line of account-centrism ends up landing in a fairly subjective place.

The cost of misjudging a real person

Does the bias carry a cost? The people in the discussion gave a very concrete answer. Asked whether these years of active enforcement had ever drawn an appeal, the reply was almost “no.” Only one or two real-estate agents and one account running outsourced PR came to ask why they'd been banned; they looked like real people, not organized behavior, so they were unbanned. Appeals being extremely rare was read as a sign that the false positive Mistaking normal content for junk and blocking it wrongs an innocent person. rate is very low. But right after, the same operations member added a more troubling admission of their own.

“I really have accidentally deleted a user, and when the user came to ask I just had to apologize … it was more like, the user had written a lot of product-introduction posts, and I assumed they were doing paid promotion, and only when they emailed me did I realize they were a real person.”

This is account-centrism's clearest failure mode: a real person who wrote a lot of product introductions gets handled as a paid-promotion account under the intuition that “the posts look like ads.” It matters not only because it was a misjudgment, but because it collides with product's belief that “people change.” If the system is willing to grant that a person may move from a rule-breaking state back to a normal one, it should also grant that a person judged to be in violation may not have been a violator to begin with. With a single real apology, this case raises a still-unanswered question: in a handling process with no clear definition, running mainly on intuition, who answers for the false positives?

This also explains why the team takes “irreversibility” so seriously. In the latter half of the same discussion, they drew a line for future automated handling: the reversible ones (quiet room, mute, freeze) can be considered for automation, with a three-day reversible window; the irreversible one (deletion) is never batched, never automated, and always kept manual. The very existence of this bottom line is itself an admission of the fact that “we do delete real people by mistake.”＊小黑屋討論逐字稿（未校訂內部討論）Deletion we can do by hand … the irreversible ones are best not batched, not automated

The spam judgment and self-censorship on the user side

The question “who is a spammer” has another dimension in the user interviews. Following the principle of informed consent, the interviewees below are all referred to by code names, and no real site name is identified.

Interviewee A is one of the few people on the site who proactively uses the report feature. She describes her criteria clearly: porn accounts, and the kind of bot-like comment that “drops a supposedly relevant line without having read the article, then, after a blank space, starts on some hospital's malpractice.” She only reports comments under her own articles, never other people's. This is a very concrete, personalized “felt sense of spam” — the reaction of a long-time creator to their own space being polluted, not a probability a model spits out. It also shows that the intuitive judgment of account-centrism lives not only on the operations side but on the user side too.

But when the same interviewee turns to the platform's most precious quality, she offers an observation that flips the word “censorship” on its head.

“The most important thing is expression without censorship! … The only thing left is my own self-censorship — I think everyone carries a censor inside their own heart.”

She even says she skips certain writing events precisely because “my self-censorship runs especially deep,” afraid of revealing too much personal information and being recognized from real life. In a space with no censorship by the platform, the strictest and most continuously operating censor is the user herself. For this whole article, this is a necessary reminder: while we argue over whether a platform should define the spammer, and whether to use account-centrism or work-centrism, the people being governed carry, all along, another set of criteria the platform can neither see nor reach.

Another interviewee, meanwhile, punctures the fragility of “quality” from the production side. Speaking of the platform's timed writing events, he bluntly says that a pace of seven pieces in eight days forces out mostly AI ghostwriting.

“What comes out in eight days is usually written by AI — I know this perfectly well about myself, I use AI to write too, and so do other citizens; you can tell, I just don't want to call it out.”

Put this line back against the earlier internal discussion, and you can see both sides facing the same phenomenon. Operations agonizes over “should AI flood-posts go in the quiet room,” unable to tell whether it's content or an ad; on the user side, AI ghostwriting has already entered the everyday creative workflow, used even by long-time users who sincerely take part in events. Part of why “who is a spammer” is so hard to answer is that creation itself has become a human-machine blend. This makes account-centrism's implicit premise — “if a real person wrote it, it isn't spam” — unstable.

The governance meaning of not defining the spammer in advance

Gather these threads together, and this team's choice turns out to be more than a technical oversight. Not writing down “who is a spammer” is a refusal to permanently define a person by a numeric line; re-evaluating every month is an assumption that people change; reserving irreversible deletion for human hands is an admission that they may misjudge a real person. These designs interlock, and together they point to the restraint a censorship-resistant platform ought to have: the object of handling should be present behavior, not a person's identity.

But it also needs saying plainly that this restraint has clear limits. The belief that “people change” lacks the matching provision to unflag, and in the end never truly lets people back in. Account-centrism and work-centrism have long gone without enough dialogue, letting the same problem grow two solutions that don't fully connect. And into the judgment of “obviously junk” too much unspoken subjectivity has crept. These problems can't be solved by writing a better spec sheet; they are governance marks left behind by a real organization repeatedly adjusting itself among scale, staffing, time, and belief.

So this article offers no standard answer to “who is a spammer.” For the policy and engineering communities, what matters more is settling a few things before building a labeling system: whether the label is permanent, whether it can be revoked, who revokes it, and whether — when the platform judges wrongly — the misjudged real person has a path back.