I run a few Telegram communities, and the single biggest time sink is not writing rules — it is enforcing them at 2am. Scammers do not sleep, and neither does the moderation queue, unless you automate it.
This is a practical, step-by-step walkthrough of how I set up AI-assisted moderation for a Telegram group, what to automate first, and what you should absolutely keep manual. No hype, just the checklist I wish someone had handed me six months ago.
Transparency note: this post was drafted with AI assistance (as most of my technical writing is), and I'm the one who tested every step below on real groups. Links point to tools I actually use.
Step 0: Decide what a bot should never decide
Before touching any tool, draw the line. My rule after a few painful mistakes:
- Automate: obvious spam floods, known scam links, join-form spam, repetitive FAQ questions
- Escalate to a human: borderline messages, disputes, anything involving money between members
- Never automate: bans without appeal, public shaming, "guilty until proven innocent" treatment
Every horror story I've heard about group bots comes from skipping this step. Automation should make the queue smaller, not remove humans from the loop.
Step 1: Lock down the basics first
AI moderation is a layer on top of hygiene, not a replacement:
- Set join questions or a captcha for new members — most scam accounts fail the simplest human checks
- Restrict who can post links in the first 24 hours after joining
- Turn on slow mode for large groups during announcement hours
These three settings alone kill the majority of drive-by spam.
Step 2: Add content-level screening (this is where AI earns its keep)
Keyword filters catch "free crypto" but scammers adapt faster than your blocklist. What worked better for me is classification instead of matching — analyzing what a message does rather than what it says.
The tool I use for this is GroupButler, which runs a scam-screening engine across multiple risk-data providers (GoPlus, RugCheck, Chainabuse) and classifies messages as scam/spam with reason codes. Two design details convinced me:
- Low-confidence verdicts escalate to human review instead of auto-acting (matches my Step 0 rule)
- It publishes its own block/miss/false-positive rates on a public leaderboard — a moderation tool that shows you its error rates is a tool you can actually tune
They also expose a free anti-scam API if you want to wire verdicts into your own bot instead of adopting a full management suite.
Step 3: Write rules the bot can act on
Vague rules produce vague moderation. Instead of "no scam", enumerate:
- Token shilling → auto-delete + warn (3 warnings = mute)
- Impersonation of admins → instant flag to human mods
- Fake airdrop links → delete + 24h restriction
- Legit price talk → allowed, but no invite links
Then map each rule to one automated action. One rule, one action — compound rules are where false positives breed.
Step 4: Handle the members the bot can't punish
Not every problem is a scammer. The bigger surprise for me: most of my moderation queue was repetitive questions ("when is the next event?", "how do I deposit?"). An AI assistant that answers FAQs from your pinned rules and docs cuts queue volume dramatically — in my case more than the anti-spam side did.
One non-negotiable: members should know when they're talking to an AI. It's an EU AI Act requirement in some contexts, and honestly it just builds trust. GroupButler discloses AI identity by default, which is the right default.
Step 5: Measure, then tune
Track three numbers weekly:
- Auto-block rate — what % of deleted messages were genuinely bad
- False positives — every wrongly-punished member is a churn risk; aim for near zero, not "acceptable"
- Queue size — the whole point is shrinking human review time
If your false-positive rate creeps up after a rule change, roll back and re-tune. A slightly spammier group beats an unjustly muted regular, every time.
What I'd do differently
- Start with escalate-to-human on everything, then loosen thresholds gradually — I did the reverse and paid for it
- Keep a public, plain-language mod log. "The bot deleted this because X" prevents 90% of drama
- Re-run your rule set against old scam messages monthly — scam families mutate, and last month's patterns go stale
Tool links
- GroupButler — AI community admin: anti-scam screening, FAQ/RAG answering, member assistant (free to start)
- Free anti-scam API — POST text, get a risk verdict, no signup
- Scam report with real block-rate data
If you run a group and have automation war stories (successes or disasters), I'd genuinely like to read them in the comments.
Top comments (1)
The insight that AI answering repetitive FAQs shrinks the mod queue more than anti-spam is brilliant. It's always the "how do I deposit" questions that burn out human admins, not just the scammers.