The Invisible Platform Janitors

Your turn

If you filtered by a threshold, where would you set it?

Even the smartest method runs into misjudgments. This time you play the moderator: try dragging the threshold. The stricter you set it, the more cleanly the junk is blocked — but legitimate content gets blocked along with it, especially the posts that just “look like junk”: political commentary, academic footnotes, and anti-scam explainers.

So a platform is better off setting a principle: when you're not sure whether a post is a problem, better to miss one piece of junk than to wrongly delete one legitimate voice.

Read more: false positives and the ethics of thresholds

Keep trying · Rounding up a spam ring

As the saying goes, where there's one cockroach, there's a dozen more…

Some junk a threshold filter simply can't catch. Click “single-post model” first, and only two obviously spammy posts get flagged — but that doesn't mean the rest are clean. Now hit “ring detection,” and the eight “translation agency” accounts that slipped through all surface at once. Each one looks like a normal post on its own, yet they share the same template, spread across many accounts, and post in a tight burst. A whole cluster of accounts moving together is what we call a Ring A ring is a group of accounts flooding the feed with the same template. Any single post scores low; you can only catch it by looking at how the whole cluster behaves. ; that's why the platform judges by “behavior patterns” instead of comparing content word by word.

Read more: how spam rings get caught, and how traditional-vs-simplified characters game the gaps

Beyond the model

Machines have limits — a model alone isn't quite enough

The quiet room

Nothing is deleted — the account or post is just moved out of public view, and can be put back at any time.

Human labeling

Wrongly removed content is retrieved and re-labeled by real people, then fed back to train the model. Ideally a human oversees the whole loop.

Retraining

Over time a model “poisons itself” and drifts further off; recalibrating it with a stronger, larger model can cut the false-kill rate to a tenth of what it was.

Community watch

A user-run “neighborhood watch” sweeps up the junk together — removing and reporting in one move, with every record kept open and transparent.

Try this next · The spectrum of measures

The “quiet room” from the toolkit is really just one of the measures a moderator can reach for. Line them all up from lightest to heaviest and it's plain to see: the heavier the measure, the harder it is to undo the cost of punishing the wrong person.

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.

Read more: a moderator's seven measures

You can also try · The home-feed ranking algorithm

Who makes the front page? And who designed it?

Platform governance isn't only about “sweeping up junk” — it's also about “who gets to go viral.” First, guess which post the algorithm will rank first on the front page, then see how two orderings — “by time” and “by algorithm” — differ. In the Matters algorithm, comments carry the most weight: what the platform most wants to lift up is content that sparks discussion, and raw reach matters less. On other social platforms it's a different story, though.

Read more: platform algorithms and the doomscroll

One more to try · How to avoid wrongful removals

Why is someone always wrongly removed?

Every post gets a score — the probability that it's junk. Legitimate content mostly lands in the lower left, junk mostly in the upper right, and in between lies a gray zone where the two overlap. Drag the dividing line and you'll find that the very same line always deletes some legitimate content on one side while letting some junk through on the other. That line is the algorithm — and there's always a chance of being wrongly removed.

So what would make it better?

Wrongly removed — how do you get it back?

Since wrongful removals can't be avoided, we have to talk about recourse. Beyond sharpening the algorithm to delete fewer good people, what matters more is whether the wrongly removed can be brought back, and whether the whole moderation system can be laid out in the open for public scrutiny. For platform governance to earn trust, we need to secure three things:

Last one to try · Censorship resistance

Delete me by mistake, and there are still thousands upon thousands of me

Beyond recourse there's another, lesser-known path: censorship resistance. The HTTP we use every day is “location addressing” — content is tied to a single server, and deleting it returns a 404. IPFS switches to “content addressing”: the same content has a unique fingerprint (CID) and lives across many nodes, so deleting one still leaves the others readable. Take it down, and compare the two fates.

Read more: how do you make digital content resilient?

The bigger question

Moderation and censorship resistance are two sides of the same coin

Censorship resistance →

A platform can block harmful content — but it can also block what it shouldn't. Distributed storage (IPFS) means a post survives even after it's deleted.

Ranking & the doomscroll →

Even when nothing gets deleted, the order you see things in was decided long ago. Deciding who gets seen is where a platform's real power lies.

When a whole domain gets blocked →

During Lunar New Year 2025, Matters' domain was mistakenly blocked by Taiwan's authorities (DNS-RPZ) for three days. Users get wrongly removed — and so do good platforms.