Rules
Part of Content moderation policy: the parts worth your attention
Content moderation policy changes: the practical version
Content moderation policy changes: what happens to material published under an old rule, how to phase a change, and the shifts that leave no document behind.
Changing a rule is not the same as writing one. A new rule arrives into a world that already holds millions of items published under the old one, reviewers trained on the old one, and users who made decisions relying on it.
Most of the damage from a policy change comes from the transition, not from the rule.
What to take away
- Decide what happens to existing material before you publish the change. Silence means somebody in a queue will decide it for you.
- Enforcement changes without any document changing, and that is the category nobody records.
- Tell the people who relied on the old rule, and give them time proportionate to what compliance costs them.
Six kinds of change
Clarification. The wording changes and nothing else does. Common, and constantly mistaken for a substantive change, which is its own cost, because people re-plan around nothing.
Six kinds of policy change
Change
- Clarification
- Wording only
- Narrowing
- Prohibited becomes permitted
- Widening
- Permitted becomes prohibited
- Reclassification
- Outcome changes
- Threshold movement
- Enforcement retuned
- External obligation
- Required by outside body
What moves
- Clarification
- Narrowing
- Widening
- Reclassification
- Threshold movement
- External obligation
Narrowing. Something previously prohibited is now permitted. Easy to communicate and easy to get wrong operationally, since existing enforcement records and strikes may relate to conduct no longer prohibited.
Widening. Something previously permitted is now not. This is the one that needs a transition plan and almost never gets one.
Reclassification. The rule is the same and the outcome attached to it changes: removal becomes reduced distribution, or a warning becomes an account action. Often more consequential than a text change, and almost never announced as a change at all.
Threshold movement. Nothing in the document changes. A detection system is retuned, a category gets a temporary sweep, a market gets a different setting. From outside this is indistinguishable from a rule change, and from inside nobody thinks a rule changed. The gap between the written rule and the enforced one is described in how the two come apart.
Implementation of an external obligation. The rule exists because a body required something of the service. It will look like a product decision, it will often be applied more broadly than the obligation requires, and its clock is set elsewhere. Which is which, and how obligations arrive, is covered in the routes regulation takes.
The retroactivity question
Decide it explicitly, write it down, and put it in the announcement.
The defensible default is that conduct is assessed against the rule in force when it occurred. That is what people plan around, it is what makes a policy something you can comply with, and it is the position an appeal will invoke. Applying a new prohibition to old material means the person had no way to comply at the time, which is a decision an operator can make and should make consciously.
Where old material has to come down for reasons that will not wait, say so plainly: this is a removal of existing material under a new standard, not a finding that anyone broke a rule. Recording it as a violation instead is how strike systems fill up with conduct that was permitted when it happened. This is also why a dated archive of the rules matters to a user, a point developed in keeping your own copy of the terms.
Phasing a widening
| Stage | What happens | Typical length |
|---|---|---|
| Announcement | The change is published with an effective date | Before anything else |
| Grace period | New material assessed under the new rule, no account consequences | Proportionate to the cost of compliance |
| Existing material | Stated explicitly: unaffected, or removable on request, or actioned | Decided at announcement |
| Full enforcement | Consequences apply | On the stated date |
| Review | The hard cases from the first weeks feed back into the text | Fixed, within a quarter |
The length of the grace period should track what compliance actually costs the affected people. A change that requires someone to re-edit a large archive is not a two-week change, and a change that requires nothing is not improved by a delay.
Telling people, in the right order
Reviewers first, with the examples, before the public announcement. A change that reaches users before it reaches the queue produces a period of inconsistent enforcement that nobody can explain afterwards.
Telling people, in order
- Reviewers first, with the examples
- Then the affected accounts, directly
- Then everyone, with what changed
Then the people it affects most, specifically. A general announcement reaches everybody equally, which means it reaches the affected minority as weakly as everyone else. If you know which accounts publish in the affected category, tell them directly. A content moderation policy checklist covers the stop conditions and the run list for each decision before a change goes live.
Then everyone, with what changed and what did not. Say plainly which of the six kinds of change this is, because "we have updated our policies" describes all six and informs nobody.
Keeping the version history
Retain previous versions, publicly if you can. The reason is not transparency in the abstract: it is that a person disputing a decision needs to show what the rule said at the time, and an operator disputing a claim needs the same thing. Both sides are worse off without it.
Where you are on the receiving end of somebody else's change, the same problem appears from the other direction, and there the only reliable record is the copy you took. The mechanics of maintaining that record, including how to supersede an entry rather than overwrite it, are in tracking a change without losing the earlier state.
Where the change is not yours to make
Some rule changes originate outside the service entirely. One country's copyright office describes its own statutory notice-and-takedown scheme at an overview of that framework, and the European Union describes its own rules for online services at the Commission's page on the Digital Services Act. Both are external regimes that services implement in their own words, in their own documents, on their own schedule. Neither describes what applies to you.
For anyone tracking changes, the announcement will not say where the rule came from. The difference matters: a rule the service chose can be argued with, and a rule the service implemented cannot be argued with there at all. What can and cannot be achieved through the service's own process is set out in the limits of an appeal.
Common questions
How much notice is enough?
Enough that someone can comply without dropping everything, which depends entirely on what compliance requires. A useful test: could a person doing this part-time meet the deadline without abandoning something else.
Do we have to announce a clarification?
Announce it as a clarification, briefly, and say that nothing changed substantively. Silent edits to a live rule are how a policy page loses credibility, and the cost of that is paid on every future change.
What about a change we are not allowed to describe?
Where an obligation genuinely constrains what can be said, say that the change exists and that the reason cannot be described, rather than presenting it as a product decision. Users notice the difference eventually and the second option is worse.
Should strikes issued under an old rule be cleared when the rule narrows?
If the conduct is now permitted, keeping a consequence for it is difficult to defend. Clearing them is operationally annoying and is the position that survives being explained.
