LLM comment detection for community platforms
Problem
Long-standing community members watch HN and similar forums get flooded by LLM-written comments run as mini bot farms to upvote posts and manipulate conversations about startups. The rate of growth is 'demoralizing', and readers who work with LLMs can spot the tone but have no tooling to verify or filter it.
Opportunity
A moderation layer / browser extension that scores comments for LLM-generation signals (stylistic, behavioral, account-history patterns) and lets communities or individual readers filter or flag synthetic engagement without heavy-handed identity requirements.
Market analysis
Demand is undeniable (HN officially banned AI-generated comments in March 2026 and Ask HN threads keep appearing), but stylistic detection of short comments is unreliable, and false-positive harm falls hardest on non-native English writers.
Market · Community moderators and engaged readers of HN, Reddit and similar forums; platforms are actively searching for tooling as AI slop floods in.
Pricing · GPTZero charges $14.99-23.99/mo consumer tiers plus custom API plans; a browser extension could run $3-5/mo, moderation APIs per-call.
Pros
- + Strong demand signal: HN updated its guidelines to ban AI-generated comments, and Reddit and Wikipedia are wrestling with the same influx.
- + Behavioral signals (account age, karma curve, upvote timing, posting patterns) are more reliable on short comments than prose style.
- + Browser-extension distribution is cheap and needs no platform cooperation.
Cons
- − Stylistic detectors perform poorly on short text; even GPTZero is documented as weak on social-media-length snippets.
- − False positives on non-native speakers and dyslexic writers are a serious harm vector, as HN moderation discussions themselves acknowledge.
- − Platforms resist third-party moderation layers; HN's own approach is human judgment plus rules, not scoring.
Existing / similar tools
Source
Hacker News (Ask HN)
The realistic product is not a detector but a triangulation dashboard: combine an LLM-tone score with account-history features (karma curve, posting times, reply latency) and let humans make the final call, the way spam filters show their evidence. Selling to platforms is a slow enterprise grind a solo builder cannot fund, so a browser extension for individual readers is the only viable wedge. The uncomfortable truth from the HN threads themselves: well-written humans keep getting called bots, so any tool that auto-hides rather than annotates will poison a community faster than the slop does.