Which Telegram comment filters to turn on

In short. There are eleven switches in the moderation settings and four are on: links, invite links, duplicates and an AI pass over the message text. The other seven are off because every one of them deletes real comments along with spam, in a ratio that depends on your community rather than on us. This page is what each one actually matches, what it gets wrong, and the order worth enabling them in — which is not the order they appear in the settings.

The default set is not a starting configuration you are expected to replace. It is the set that survives contact with a normal comment section without anyone complaining, and for a lot of channels it is also the finishing configuration.

The four that are already on

Links. Any message containing a URL is deleted. This is the single highest-yield filter in a comment section, because almost all comment spam is a link with words around it. It also has an allowlist — put your own domains in it before you wonder why your own links vanish.

Invite links. t.me/joinchat/… and t.me/+… specifically. Separate from the links filter because “come to my channel” is a different problem from an advertising URL, and people who turn off the general link filter usually still want this one.

Duplicates. Three identical messages inside five minutes are treated as flooding: deleted and muted for five minutes.

That window matters more than it looks. The filter originally ignored it, which meant three identical messages ever — someone who wrote “thanks” under three different posts across two months got muted. That is fixed, and it is a useful illustration of how a filter that seems obviously correct can be wrong in a way you only see in production.

The AI pass. The message text goes to a model that answers spam or ok. It exists because pattern matching cannot catch the message that has no link, no phone number and no forbidden word — the one that just reads like an advertisement. It costs you nothing and is not gated by plan.

It is capped at 2,000 checks per workspace per day. That ceiling is not about your plan; it is because a raid maximises the cost of classifying it, and past the ceiling the rule-based filters keep working alone.

The seven that are off, and what each costs you

Roughly in the order worth turning them on.

A stop-word list. Off because it is empty — nobody can guess your words. This is the highest-value switch here precisely because you write it: the crypto ticker, the competitor’s name, the phrase your particular spammers use. Costs you nothing you did not choose.

Phone numbers. Deletes messages containing something phone-shaped. Safe for most channels, wrong for anything where people legitimately swap contacts — a classifieds or services community will lose real messages to this one.

Forwards. Deletes forwarded messages. Effective against the copy-paste crowd, and it also deletes the member who forwards a genuinely relevant post from another channel. Turn it on if your comments are for discussion, leave it off if they are for sharing.

Mentions. Deletes messages containing @username. Read this one carefully: it fires on any mention, including a member tagging another member mid-conversation. It also used to fire on the @gmail inside an email address and delete the message as spam. That is fixed, but the incident describes the filter accurately — it is a blunt instrument that happens to catch a lot of “DM me @someone” spam.

New-member restrictions. For a set number of hours after joining — 24 by default — a member cannot post links or forwards. Media and stickers are separately switchable and off. This is the gentlest of the seven, because it expires by itself and hits exactly the population that spams.

Night mode. Between two hours you set, links and media are blocked. Useful if your raids are nocturnal and you are not. Useless, and mildly insulting to your audience, if they are not.

Reputation gating. The harshest, and the one to understand before enabling. It scores a member as days since joining minus three per warning, and deletes their links while that score is below a threshold — ten by default. A perfectly genuine member who joined last week cannot share a link. Under attack that is exactly what you want; as a standing rule it quietly makes your comment section unwelcoming to everyone new.

What happens to a repeat offender

The escalation is separate from the filters and applies to all of them: three violations mute for an hour, six ban.

The thing worth noticing is what it does for a first-timer. Somebody who pastes one link and gets it deleted has lost a message, not their access. That is deliberate — the cost of a false positive on the first offence is small, which is what makes the more aggressive filters survivable at all.

The order to turn them on

Not all at once, and not because it is risky — because you will not know which one caused a complaint.

1. Live with the defaults for a week. Read the log. Most channels stop here.
2. Add the stop-word list once you have seen the same spam phrase three times. You are encoding an observation, not guessing.
3. Add phone numbers unless people legitimately share contacts.
4. Add new-member restrictions if the spam arrives from accounts that joined hours ago.
5. Forwards and mentions only if the log says so. These are the two that annoy real members.
6. Night mode and reputation are raid responses. Turn them on during, off after.

If… then…

If…	Then…
You just connected the channel	Change nothing for a week; read the log first
Your own links keep vanishing	Add your domains to the link allowlist
Spam has no links and no keywords	The AI pass is already handling it
The same phrase keeps arriving	Stop-word list — you have the observation now
Your community swaps phone numbers	Leave the phone filter off, whatever the spam looks like
People forward relevant posts to each other	Leave forwards off; it will cost you more than it catches
You are being raided right now	Night mode and reputation gating, then turn them back off
A new member complains they cannot post a link	That is reputation gating, and it is working as designed

Where to start

Open the moderation log before you open the settings. It records what was deleted and which filter did it, and a week of that tells you more than any recommendation on this page — including whether you have a spam problem worth configuring at all.

If the log is nearly empty with the four defaults running, you are done. That is a legitimate outcome, and a comment section that deletes almost nothing is the goal rather than a sign the filters are not working.

---

For the join gate that runs before any of this, see protecting a Telegram channel from spam. For how the filters behave message by message, see Telegram comment moderation. To set them up, start at the protection page.