How Telm spam detection works: layers and verdicts

1One bot, many layers

Telm does not rely on a single trick. Think of it like security at an airport: a message passes through several checkpoints, one after another, and can be cleared or stopped at any of them.

That layered design is what makes protection both reliable and fast. Obvious cases are settled at once, and only the genuinely ambiguous messages are looked at more closely — which is why even a busy group of tens of thousands of members is moderated in real time without ever feeling laggy.

• Global spammer check — known abusers are stopped at the entrance, before their first message.
• Rule and pattern detectors — a large library that recognises spam schemes on sight.
• Repeat-spam memory — a wave of identical or look-alike spam is caught at once.
• Our own ML model — trained on real spam in many languages to catch campaigns it has never seen.
• AI classifier — reads meaning, for the rare message that is genuinely borderline.
• Self-learning — your decisions make every layer sharper for your group.

2The journey of a single message

Here is what happens when someone posts in your group. The message is checked against layer after layer, and most messages are cleared or caught long before the end.

Admins, the group owner and trusted whitelisted members are never moderated — their messages skip these checks entirely, which is why you never have to worry about the bot touching your own team.

• Entry check — a known spammer is stopped at the door by the global spammer database before their first message (see CAS blocklist and the global spam database).
• Rules and patterns — many detectors add up a spam score. No single word bans anyone; the decision comes from the sum of the signals.
• Repeat-spam memory — if the same or near-identical text was just judged in your group, that answer is reused, so a flood of copies is handled in one go.
• Our ML model — scores how spammy the wording looks; confident spam and confident clean text are settled here, and the middle moves on.
• AI layer — a language model reads the meaning of a still-ambiguous message (see AI moderation).
• Images and audio — perceptual hashing, OCR text recognition, QR-code detection and voice transcription feed their text back into the very same checks.
• Verdict and action — pass, delete, mute, ban or send to admin review. Everything is logged with the reason in the moderation journal.

3What each layer actually does

It helps to know what job each checkpoint is really doing, because they catch very different kinds of spam. Together they cover each other's blind spots.

• Rule and pattern detectors — hand-built detectors that recognise spam schemes rather than single words: move-to-DM lures, hidden ads, external-channel promotion, structural tricks and obfuscation. They work on normalised text, so mixed alphabets and inserted symbols do not fool them.
• Repeat-spam memory — recognises identical and near-identical messages (a swapped phone number, a different handle) and reuses the earlier verdict, so coordinated floods are shut down the moment they start.
• Our own ML model — a machine-learning classifier trained on a huge collection of confirmed spam plus clean speech in many languages. It generalises to campaigns it has never seen, including obfuscated and translated variants.
• AI classifier — a language model reserved for the genuine grey zone. It reads intent, so it can tell a real question from a disguised pitch.
• Image and audio understanding — OCR pulls text out of banners and screenshots, perceptual hashing matches known spam pictures, QR codes are decoded, and voice and video notes are transcribed. All of that recovered text re-enters the same rule, model and AI checks.

4Scores, thresholds and actions

Instead of a blunt yes/no, each message accumulates a spam score from many signals. The higher the score, the stronger the action — and a single unlucky word can never reach a ban on its own, because the contribution of keywords is capped well below the punishment thresholds.

This is the built-in safeguard against false bans: a verdict is never handed down by words alone, only by the full picture. The strictness of these thresholds is something you control through moderation strictness levels and protection presets.

• Very high score — automatic removal, and for clear abuse a mute or ban.
• Medium score — sent to an admin for review rather than acted on blindly.
• Low score, nothing suspicious — the message is released immediately.
• New or untrusted members face slightly firmer thresholds than long-standing trusted regulars, so trusted members almost never get caught.
• Every action is reversible and appealable, and each decision records which layer decided and why (see understanding moderation decisions).

5What it catches

Rather than matching stop-words, the engine recognises spam schemes by their whole pattern, including obfuscated, image-based and spoken variants. These are the families it is built to stop.

• Move-to-DM lures — invitations to continue in private messages, even when phrased like a normal remark.
• Hidden ads and third-party channel promotion, including forwards from other channels (see controlling forwards and channels).
• Coordinated campaigns — the same text surfacing across many groups, or from several accounts inside one group in a short window.
• Scam and phishing patterns — fake giveaways, investment bait and impersonation of admins (see how scam detection works).
• Image spam — casino banners, QR codes and text baked into pictures.
• Voice and video-note spam — pitches and links read aloud to dodge text-only bots.
• Obfuscated text — mixed alphabets, inserted symbols and look-alike characters.

6How false bans are prevented

The biggest fear with any anti-spam bot is that it removes a real member by mistake. Telm is designed so this is both rare and always recoverable, because every decision is transparent and reversible.

Several safeguards work at once, and all of them put human judgement above the machine.

• Score capping — keywords alone can never reach a ban; it takes the combined weight of many independent signals.
• Admins and trusted members are never moderated, and you can add more names to your whitelist and trust list.
• The medium-confidence zone goes to admin review instead of an automatic punishment.
• Every action is logged with its reason and can be appealed — approved appeals clear similar messages too (see how appeals work).
• An optional personal ML model learns your community's normal speech and vetoes false alarms that do not fit it — it can only clear messages, never punish.

7It learns from you

The system is not frozen. Your moderation decisions and approved appeals continuously teach it, so it becomes more accurate for your specific community over time.

Every one of these feedback loops sits below human judgement — when you approve an appeal or mark a false positive, that verdict overrides the machine and clears similar messages too.

• Your corrections keep improving the ML model, so detection sharpens for the kind of messages your group actually sees.
• Marking one message as a false positive instantly clears the look-alike messages around it.
• New detectors are proven on real traffic before they are allowed to act, and pulled back if they start to slip.
• A member's track record adjusts how much slack they get: a history of clean messages earns a little leeway, a punishment tightens it.
• A personal ML model can be trained on your own chat to clear false alarms on your community's normal way of talking.

8Staying in control

You decide how strict the bot is. Two settings matter most when you start, and everything the bot does is visible and adjustable.

Monitoring Mode lets you watch what the bot would do without it actually punishing anyone — the safest way to build trust before switching on real enforcement.

• Monitoring Mode — observe only, no punishments, everything is logged for review.
• Strictness levels and presets — pick a ready profile or fine-tune thresholds for your community.
• Relaxed rules — name specific detectors that must not score in your group (critical protections cannot be relaxed).
• Allow sales and crypto community mode — soften triggers for communities where trading talk is normal.
• Per-group toggles for CAS, the global spam database, AI moderation, voice moderation and the personal model.

9Best practices and common mistakes

A few habits get you the most protection with the fewest false alarms. Most problems new admins hit come down to skipping one of these.

• Do start in Monitoring Mode and read the journal for a few days before enforcing.
• Do whitelist your own team and known partners so their links are never questioned.
• Do enable the global spammer check straight away — it removes attacks before you ever see them.
• Do act on appeals and false positives; every correction makes your group's protection sharper.
• Do not disable whole protections to fix one false alarm — relax the single rule involved or whitelist the member instead.
• Do not expect brand-new spammers to be on any external list; that is exactly what the rules, model and AI layers are for.