AI moderation in Telm: the LLM classifier explained

1What AI moderation does

AI moderation switches on a language model (LLM) that reads the meaning of a message rather than scanning it for banned words. It is the layer that tells apart where can I buy coffee from write to me privately, I will teach you to earn.

It sits on top of the rules and our own ML model. Most messages never reach it — it is reserved for the genuinely ambiguous cases the other layers could not decide on their own.

2Why intent beats keywords

Keyword filters have a ceiling. A careful spammer never writes the banned word — they hint, they misspell, they wrap the pitch in friendly small talk. A rule engine sees nothing wrong; a human reading the sentence sees the intent instantly.

That is the gap AI moderation fills. It judges what a message is trying to do, so it can separate a real opinion about crypto from a disguised investment scam, or a genuine question from a move-to-DM lure dressed as curiosity.

• Understands paraphrase and misspelling instead of exact strings.
• Reads the whole sentence in context, not isolated words.
• Tells legitimate topic discussion from a hidden pitch on the same subject.
• Reduces false positives because it grasps meaning rather than guessing from words.

3When it is worth enabling

Telm's other layers already stop the bulk of spam. AI moderation is the extra edge for communities where spammers write carefully, without obvious keywords.

• Your group attracts disguised promotions and move-to-DM lures phrased like normal chat.
• You see spam that reads as a genuine question or opinion at a glance.
• You want the bot to understand your topic — for example, to tell legitimate crypto discussion from a hidden scam.

4How it fits the pipeline

AI moderation is not a separate bot — it is one more layer on the detection ladder. It only engages after the other layers have narrowed a message down to a real grey area.

Anything still unclear after the AI looks at it is handed to human review rather than quietly slipping through, so a busy group never trades accuracy for speed.

5Accuracy, memory and review

AI moderation is tuned to be careful, not trigger-happy. A confident spam verdict acts; anything it is unsure about is routed to admin review rather than punished, so the AI never removes a member on a hunch.

Its confident verdicts are remembered, so the same or a look-alike message is handled instantly the next time without going back to the AI. That is what keeps the feature both accurate and quick over time.

• Confident AI verdicts are remembered and reused for identical or near-identical text.
• Uncertain cases become an admin-review item, visible in the moderation journal.
• During a sudden flood, anything the bot cannot immediately clear is sent to review, never passed through unchecked.
• Your approvals and rejections feed back into the pipeline the same way as any other decision.

6Quotas by plan

Because AI runs last and only on the grey zone, the vast majority of messages are settled by the other layers before the AI is ever called — so it never gets in the way of a busy chat.

There is nothing to pay per message: the AI check is part of your plan, with a daily quota of 100 checks on Basic, 1,000 on Pro and 10,000 on Business — the same quota also covers the public /spam/check API. If a sudden influx ever pushes more borderline messages than the quota allows, the rule engine, CAS and the global database keep protecting the group.

7Best practices

A short break-in period gives you confidence before you let AI act on its own.

• Turn it on with Monitoring Mode first and read the AI verdicts in the journal for a few days.
• Keep your whitelist and trust list current so partners and regulars are never questioned.
• Pair it with the personal ML model if your community has a very distinctive way of talking.
• Do not treat AI as a replacement for rules — it is the top of the ladder, and the other layers still do most of the work.