How It Works
Rule-Based Filters
Fast regex patterns and heuristics catch obvious spam instantly.
- Regex pattern matching
- Known spam phrase detection
- URL blacklist checking
- Character encoding analysis
Local AI Model
On-device machine learning model analyzes message patterns, user behavior, and context.
- Behavioral pattern analysis
- Message sequence detection
- User reputation scoring
- Context-aware classification
Cloud AI Analysis
For edge cases, our advanced cloud LLM provides nuanced understanding.
- Natural language understanding
- Intent classification
- Multi-language support
- Continuous learning
What It Catches
Why 3 Levels?
Single-layer detection systems face a fundamental tradeoff. Our 3-level approach solves this.
- Level 1 handles most spam with near-zero latency.
- Level 2 catches additional spam with behavioral analysis.
- Level 3 handles the most sophisticated remaining spam.
Why rules alone stop being enough
| Challenge | Keyword rules alone | AI analysis |
|---|---|---|
| Brand-new spam | Misses anything that wasn't explicitly listed in advance | Judges meaning and intent, so a scam it has never seen still reads like a scam |
| Disguised text | Defeated by v.i.p sp@cing, homoglyphs and text baked into pictures | Reads through the disguise the way a human does — including text inside images via OCR |
| Context | Judges every message in isolation, so borderline cases become coin flips | Weighs the message together with account context — a newcomer dropping contact info reads differently from a regular sharing a link |
| Maintenance | Someone has to keep the keyword lists fresh forever | Keeps learning from decisions and cross-group signals without manual list curation |
It keeps learning — on real traffic
Our own trained model
A dedicated spam model trained on real Telegram traffic catches 97.9% of spam with just 0.04% false positives — and it answers in milliseconds, before heavier AI even needs to run.
Verdict memory
Confirmed verdicts are remembered. When a near-duplicate of known spam shows up — even reworded or padded with emoji — it's blocked instantly, with no new AI call and no delay.
Cross-group protection
A spammer caught in one community is known to every community. Detection signals are shared across groups, so a campaign that failed once doesn't get a fresh start next door.
Test mode for admins
Flip on test mode and the bot moderates the admins' own messages as if they were regular members. Post borderline content yourself and see exactly which rules fire — without punishing anyone.
How 3-level detection decides on a message
Every message runs through the levels in order, cheapest first. Level 1 applies regex patterns, known spam phrases, URL blacklist lookups and character-encoding checks — obvious spam is caught here with no model call at all. Whatever is not clearly resolved passes to Level 2, a text classifier running on our own servers that scores the message on content, context and the sender's recent behaviour without any per-message external API request. Only the genuinely borderline cases — a small slice of traffic — escalate to Level 3, a cloud LLM that reads the message for intent and disguised phrasing. The pipeline stops as soon as a level is confident.
Ordering the checks this way keeps the common case fast and the expensive case rare, so a busy group is not waiting on an LLM for every line. The threshold between "act" and "leave alone" is yours to tune, and it is worth doing: a stricter setting catches more borderline promotion but will occasionally flag an enthusiastic real user, while a looser one lets more through. Because every decision — which level fired, the score, the intended action — is written to the log, you are never guessing why a message was removed, and you can start in Monitoring Mode to watch the pipeline before it touches anyone.
Frequently asked questions
How accurate is the AI spam detection?
The layered pipeline — rules, our own trained model, and an AI classifier for the hard cases — catches about 97.9% of spam with roughly 0.04% false positives on real traffic. And you don't have to take our word for it: the bot starts in Monitoring Mode, so you can watch its decisions on your own group before it acts on anything.
Does it work in my language?
Yes. Detection is multilingual, including CJK scripts, transliterated text, and mixed-language messages. OCR also reads text inside images, so spam hidden in pictures is caught the same way.
What happens when the AI gets it wrong?
The punished member gets a polite notice in their own language and a private appeal link. An AI arbiter reviews the case with the group's context; clear mistakes are reversed automatically in seconds and the member's reputation is restored.
Do I need to configure anything?
No. Detection works out of the box with sensible defaults. If you want, you can tune the strictness, add your own keyword rules, and use test mode to feel out the thresholds on your own messages first.
Where teams use this
Real community types where this feature does the heavy lifting.
Protect your crypto group from scammers, fake airdrops, and phishing links — including scams hidden inside images.
Professional moderation for company groups and client communities. Strong protection with detailed audit logs.
Protect your fan base from scammers impersonating you. Build a safe space for your community to connect.
Related Features
Explore more ways Telm can help your community