diff --git a/proposals/3725-content-warnings.md b/proposals/3725-content-warnings.md
new file mode 100644
index 000000000..17215501b
--- /dev/null
+++ b/proposals/3725-content-warnings.md
@@ -0,0 +1,86 @@
+# MSC3725: Content warnings
+
+[MSC2010](2010-spoilers.md) made it possible to tag parts of messages as
+spoilers, so that receiving clients can hide content that the user might not
+want to see. However, the latter proposal only works with textual messages,
+which are just one of the many event types that might require some kind of
+content warning (images, videos, files, etc.). This MSC proposes a more general
+content warning system to fill that gap, using
+[Extensible Events](https://github.com/matrix-org/matrix-doc/pull/1767) to
+encompass all possible event types.
+
+As with textual spoilers, there are a variety of reasons that an event might
+warrant a content warning, for example if it could spoil a story, or if some
+recipients might find the content objectionable or disturbing. By providing a
+content warning, clients can then take measures to require consent from the user
+before displaying the content.
+
+## Proposal
+
+Content warnings are represented by a new event content type
+`m.content_warning`, which looks like this:
+
+```json
+{
+  "type": "m.image",
+  "content": {
+    "m.text": "Upload: screenshot.png (300 KB)",
+    "m.file": {
+      "url": "mxc://example.org/abcdef",
+      "name": "screenshot.png",
+      "mimetype": "image/png",
+      "size": 300000
+    },
+    "m.image": {
+      "width": 640,
+      "height": 480
+    },
+    "m.content_warning": {
+      "type": "m.spoiler",
+      "description": "Contains spoilers for chapter 6"
+    }
+  }
+}
+```
+
+The `type` field is optional, and represents the general, machine-readable
+reason for the content warning. The following types are provided:
+
+- `m.spoiler` for spoilers
+- `m.nsfw` for NSFW content
+- `m.graphic` for graphic or disturbing content
+- `m.medical` for e.g. epilepsy warnings
+
+The `description` field is also optional, and represents a more specific,
+human-readable description of the content warning.
+
+## Potential issues
+
+A result of this proposal is that textual `m.message`s now have two different
+ways to do spoilers: the `data-mx-spoiler` attribute, and the
+`m.content_warning` event type. While this may seem a bit redundant, the two
+methods are necessarily complementary, since the `data-mx-spoiler` attribute
+provides finer-grained control over which parts of a message are hidden, while
+the `m.content_warning` event type sits above it at the event level.
+
+## Alternatives
+
+In the case of images, an alternative solution, which some people currently use,
+is to embed images inline in `m.text` messages, and then tag it using the
+existing mechanism for textual spoilers. This is a pretty hacky workaround,
+since it does not support other types of media, nor does it support encryption.
+It also loses the semantics of standalone `m.image` events, which makes it
+difficult for clients to render image spoilers differently from regular textual
+spoilers.
+
+## Security considerations
+
+None that the author is aware of.
+
+## Unstable prefix
+
+Clients wishing to experimentally implement this proposal may do so by replacing
+the `m.content_warning` event type with `town.robin.msc3725.content_warning`,
+and replacing the content warning types `m.spoiler`, `m.nsfw`, `m.graphic`, and
+`m.medical` with `town.robin.msc3725.spoiler`, `town.robin.msc3725.nsfw`,
+`town.robin.msc3725.graphic`, and `town.robin.msc3725.medical` respectively.