Hypothesis: there exist one or more kinds of "objective spam"

Hypothesis: there exist some kind/s of spam in Status chat that can be, without controversy, automatically mitigated with minimal human intervention and monitoring. This kind of spam can be identified objectively by its content, i.e. it’s not a matter of participants’ tastes or inappropriateness relative to chat topic. More concretely, messages comprising this kind of spam can be measured, compared to messages not of this kind, to have a significant +/- effect on the entropy of a chat as determined and averaged over rolling intervals (10s, 30s, 1m, 2m, 6m, 18m, …).

Challenge: supposing such “objective spam” exists, implement an algorithm (and/or employ ML) that can flag it. Because the measurement is objective, the filter can be open source without concern that its being so defeats the purpose (see below). Status could recommend, as a best practice, that relay and history (mailserver) node implementers include such a filter to auto-block routing of objective spam. The filter ought to be tunable and swappable, and a node could provide an HTTP endpoint that shows a log (updated hourly) of the last 1000 messages (content only) it filtered as objective spam in public chats. Periodic or spot human review of the logs could be done to check that messages aren’t being inappropriately filtered; if so, the filter could be adjusted accordingly.

Corollary hypothesis: while not addressing all kinds of spam and moderation concerns, auto-blocking of objective spam will be an important base-layer of a multi-layered set of infrastructure-side (albeit decentralized, i.e. relay/history nodes) and client-side mechanisms for improving the Status chat experience and promoting the health of the messaging network. Insofar as the filter mechanism is open source, determined spammers will find a way to work around it (see the Conventional solutions… section of 1973). However, if the measure of objective spam is grounded on e.g. Δ-entropy and is effective (as hypothesized above) then spammer workarounds likely encompass an altogether different category of spam, to be addressed by other mechanisms, while the objective spam filter retains its beneficial effects.

4 Likes

Possibly related, reading now:

 Measurement and Classification of Humans and Bots in Internet Chat  [PDF] (2008)

1 Like

The traditional algorithm in “ML” to fight spam is called Naive Bayesian. I put “ML” in quote because it’s so old that it’s not as fancy as Transformers (the state-of-the-art language processing models).

It’s very cheap to run but the main issue is updating the training to address new forms of spam.