The discussion revolves around the frequent failure of the Akismet filter to accurately detect spam on a Discourse forum, particularly with technical posts and obvious spam. mbauman notes a 94% approval rate for the system user, indicating many false positives, and expresses frustration with Akismet missing spam (like gambling links) while flagging legitimate posts. Despite disabling Akismet and setting skip_akismet_trust_level to 0, flagged posts persist. RGJ clarifies that posts are only checked if both trust level and post count conditions are met, but a hardcoded rule always checks the first post of a TL1 user. justin suggests that training Akismet through the review queue can improve its performance. ganncamp shares similar issues and inquires about the approval rate. mbauman shows where the approval rate is displayed and questions if Akismet’s effectiveness has declined. A potential fix is referenced in a GitHub pull request 52, and mbauman confirms the patch has been applied, leading to the disabling of Akismet except for user bio checks, which remain reliable.
Our system user has the resounding approval rate of 94% . The akismet filter just completely misses on a lot of technical posts like this one — that’s a post by a TL1 user who has been around for 8 months with no outbound links. At the same time, it’s missing obvious spam with direct links to gambling sites.
We’ve tried disabling Akismet because it just creates extra work: we get far far more false positives than true positives. But that doesn’t seem to work; new topics like the above are still getting flagged, even though we’ve set “skip akismet trust level” to be 0.
Now that I’m looking into this I see we still have the default 50-post skip setting; I’m unsure how those settings interact. Is there an implicit “AND” between those two settings?
I can’t speak to the technicals of our implementation of Akismet, but I do believe we send feedback to the algorithm when you select a flag option. As you verify spam and ham posts through the review queue, you’re helping train Akismet.
@mbauman how do you find the system user approval rate? I’m sure our number would be worse. I’m not sure I remember ever having an Askimet-flagged post actually be spam. And BTW our settings are at this & still seeing posts being flagged (I’m about to adjust the trust level to match yours.)
Just to finish out this topic, the above extension patch has now propagated to our instance. We’ve disabled Akismet (except user bio checking, which is the only thing that actually seems to work with any level of reliability) with the following settings:
(It probably had propagated earlier, but I had missed it since the keyword akismet does not appear in the setting itself so simply searching akismet does not bring it up).