Comment by graypegg

On the LLM idea, if you could group reports by issue (by parsing the user provided input and whatever context you save from the page screenshot into some embedding) and then only escalate things when several different IPs have reported a similar thing within X amount of time, I think you could handle two birds with one stone. Limits how annoying spammers can be, and also makes the good reports easier to understand since a few bug reports combined should make a better whole.

I however wouldn't shorten/transform reports with an LLM, or make spammy reports inaccessible. Just doing the semantic grouping for escalation. It's true you're getting free work from your users, and the human factor is pretty important here, even if an LLM might sometimes misinterpret it.