Configurability for spam invite detection #422
Description
Describe the problem
Sable has a builtin mechanism to detect spam invites, which classifies invites as spam if they are either sent by a user that is banned in any known room, or if the invite contains text on a "bad words" list. Both of these mechanisms have a high false positive rate for me, making the feature not very useful. It is also a problem when my own legitimate invites (or legitimate invites sent by my friends) get miscategorized as spam by other people's cinny/sable clients.
The hardcoded bad words list, in particular, is very questionable. The full list is here. It mostly contains a bunch of swear words (including some that are usually considered extremely mild), words related to sexuality, and some slurs. I am in legitimate rooms with swear words, sexuality-related words, and (reclaimed) slurs in their names. (also, amusingly, most of the slurs that I actually get in spam invites are anti-trans slurs that are not on the list).
Separate from the direct concern of false positives, it bothers me to be using software that encodes the value "sex is bad" into it's behavior. I think this is a value that does a lot of harm to the world and to people I care about, and I would like to avoid propagating it if possible.
Describe the solution you'd like
The simplest thing is adding an option to turn off spam invite filtering entirely. If that option was available, I would be satisfied just disabling it.
If we wanted to make the spam invite filtering feature more useful though, there are several changes that could be made:
- allow enabling/disabling the keyword filter and the banned user filter independently
- allow restricting which rooms are checked for bans
- allow checking a policy list for bans (ie, I would be fine marking invites from any user that is banned on CME as spam)
- allow user-configurable keyword lists rather than having the keyword list hardcoded
- produce a new hardcoded list that is more representative of the spam invites that actually exist on matrix, without as many false positives
Alternatives considered
No response
Additional context
I created a similar issue in upstream cinny back in May after the invite filtering feature was introduced. I mostly summarized the content of the original issue here (I hope, with a less angry tone).