AI spam bot says it is not spam but scan log says it is spam

Contribute Bug

AI-generated summary

J-Ha_Hasegawa reported an issue with Discourse AI spam detection: a clearly spammy post was not blocked despite the scan log flagging it as SPAM. The AI test feature incorrectly returned “Not spam” — a bug later confirmed by Roman, who identified that the Spam Detector persona’s system prompt was confusing Claude models. A fix was submitted via GitHub PR #34442 to clarify response formatting, resolving the test output discrepancy J-Ha_Hasegawa. However, the post still wasn’t blocked during live posting. After investigation, Roman explained that AI spam detection skips posts from users with Trust Level 2+, staff, bots, or those with >3 regular posts. J-Ha_Hasegawa realized they had posted as an admin (TL4), which bypasses spam checks. Switching to a TL0 test user resolved the issue — the spam post was correctly flagged and unlisted by the discourse_ai_spam user. The thread concludes with gratitude for the prompt, helpful support and confirmation that the system works as intended when used with appropriate user permissions.

1

I’ve enabled the Discourse AI spam handling on our forum. I’ve set up Claude Sonnet 4 with an API key and selected the Spam detector persona.

I did a test post that is clearly spam. Nothing subtle about it.

It was not blocked and was posted immediately.

When I gave the post URL to the spam bot using the test feature, the result says Not spam, but in the Scan log it says: SPAM - This is a clear promotional advertisement…

CleanShot 2025-08-19 at 17.05.24@2x
CleanShot 2025-08-19 at 17.05.24@2x1298×1566 223 KB

My expectation would be that the result would be SPAM, matching the Scan log declaration of SPAM. And that this would then queue up the post for review by admins and moderators, for example.

Might anyone be able to share what I’m missing? I’m no expert – so am open to any guidance!

Thank you!

2

What is the trust level of the user who posted? AI Spam will skip posts from TL2+ users.

Regarding the test, there’s a bug in the test code and it should say “spam.” I’ll work on a fix.

4 Likes
4

Thank you for your reply!

The user I used to post is Trust Level new user

Any thoughts on why the post made it through?

I appreciate your help!

5

This will fix both the test and the post not getting flagged:

https://github.com/discourse/discourse/pull/34442

The Spam detector Persona system prompt was confusing Claude models. The change makes the expected response format instructions more explicit.

4 Likes
What is stopping you from trying out Discourse AI?
6

Ah, fantastic! The test feature is working as expected.

I am wondering if you might be able to help with why the AI Spam feature is still not blocking a spammy post from being immediately posted? I sent the post to the AI Spam test and it is flagging it as spam - but it was posted.

Am I missing a connecting piece perhaps? Thank you so much for your help with this!

CleanShot 2025-08-21 at 18.30.15@2x
CleanShot 2025-08-21 at 18.30.15@2x2494×1158 290 KB

1 Like
7

Are you an admin, or higher TL? If yes, then you perhaps would try to use some low TL test user instead.

1 Like
8

We skip a post when:

  • The author’s trust level is greater than TL1.
  • The post belongs to a private message topic.
  • The author is a bot.
  • The author is staff (moderator/admin).
  • The author has already made more than 3 posts in regular (non-PM) topics.
  • The post has already been scanned 3 or more times.

If the test is working, I’m confident it has to be because one of the above.

1 Like
9

Ahhh yes! Thank you for your patient and helpful replies!

I posted with my admin user instead of my trust level 0 user. :woman_facepalming:

It’s working! I love the way the discourse_ai_spam user shows up as the user who flagged and unlisted the post.

CleanShot 2025-08-22 at 08.27.35@2x
CleanShot 2025-08-22 at 08.27.35@2x1716×1544 262 KB

Thank you again for your quick and generous help with this!

3 Likes