Feature request: AI-powered “meaningless reply” filter with user-selectable levels

Feature
1

Hi folks,

I’d like to propose an AI-assisted filter for low-signal / meaningless replies (“water replies”, generic spammy comments) inside topics.

The problem

In many English-language communities, topics often get flooded with short, low-effort replies like:

  • “Thanks / nice / good post”
  • “I agree”
  • “Following”
  • “Up”
  • generic compliments with no new information

These replies add noise, increase scrolling, and reduce the reading experience for everyone—especially in long threads.

Moderators can fight this with rules and manual actions, but it doesn’t scale well. And purely regex-based filtering tends to have high false positives.

What I built / current prototype

I’m currently experimenting with a plugin-like approach:

  1. Regex pre-filter to catch the most obvious short / evasive replies
  2. Then pass the remaining candidates to an LLM for classification
  3. The UI must be transparent: show how many replies were hidden
  4. Hidden replies should be viewable on demand (“Show hidden replies”)
  5. (Optional) For hidden content, run another LLM step to extract 2–3 representative sentences from the hidden replies, so readers can quickly see “what was hidden” without expanding everything.

Even with regex only, the reading experience improves a lot (less scrolling), but the false positive rate is too high—so I believe LLM-based judgement is necessary.

Proposed product behavior (user-facing)

Inside each topic, readers would see something like:

  • “12 replies hidden (Low-signal filter: Medium)”
  • Buttons: Show hidden replies / Change filter level

And provide several levels, for example:

  • Off: show everything
  • Low: hide obvious junk (very high precision)
  • Medium: hide common low-signal replies
  • High: aggressive filtering (user opts in)

Importantly:

  • This should be per-user preference (and maybe also configurable per-category/site default).
  • The system must remain transparent and reversible: nothing is “deleted”, only hidden by default.

Why this fits Discourse (especially now)

Discourse already has multiple AI-related features and the Discourse AI ecosystem is growing. I think an in-topic reply “cleaner” is one of the most practical, high-impact uses of LLMs for community UX.

It’s not exactly “spam detection” (which is usually account-level). This is more about topic-level reading quality.

Questions for maintainers / community

  1. Does Discourse already have plans for a reply quality / low-signal filter?
  2. Would it make sense to build this as an extension to Discourse AI, or as a separate plugin?
  3. What’s the best way to implement the UI/UX so it’s transparent and doesn’t confuse users?
  4. Any concerns about moderation policy, trust levels, or edge cases (e.g., short but valuable replies like “Solved”, “+1 with a link”, etc.)?

If this direction makes sense, I’m happy to share more details (regex rules, UI mock, prompt ideas) and potentially work on a PR/plugin.

Thanks!

1 Like