Back
NoteApr ’232 min read

Bluesky is designing moderation like a system

Bluesky’s plan for composable moderation separates labelling content from deciding what to do about it. That’s how a good design system works, and it puts the hard part in the defaults.

Typeface

Last week Bluesky published its plan for moderation. The short version is in one sentence: “Moderation should be a composable, customizable piece that can be layered into your experience.”

The mechanics are what interest me. Anyone can create labels and apply them to content. People subscribe to the label sets they trust, much like shared block lists. Each label can lead to one of three treatments: hide, warn or show. Underneath all of it, automated filtering removes content that’s illegal.

That’s the structure of a good design system, applied to speech.

Separate the label from the treatment, and the same decision can be made once and applied in many different ways. One person hides everything labelled as graphic. Another sees it behind a warning. A third sees it plainly. Nobody has to relabel anything, and the people doing the labelling don’t have to agree with each other on what to do about it.

On Android I design for screens read from across a room as well as up close, and the approach I trust most has the same split. The content says what matters in it, and the system decides how much of it to show. Splitting what something is from how it’s shown lets a system serve situations its authors never drew.

The risk in this design is that it’s honest about choice and most people won’t make one. Nearly everyone will use whatever label sets and treatments they start with. So the most consequential moderation decision on Bluesky will be the default subscription a new account gets on day one.

The other risk is in who labels. Anyone can make a label set, which means anyone can make a hostile one. Subscriptions limit the damage, since a label only matters to people who chose it. But a large label set run carelessly could still shape what a lot of people see.

I’d still rather see moderation built this way than as one opaque rulebook. It can be inspected: a hidden post can say which label hid it and who applied it. And it makes the default a published, arguable choice, which is better than one nobody can see.