AI Auto-Mod

Standard Auto-Moderation matches patterns. AI Auto-Moderation reads the message the way a moderator would: it understands sarcasm, coded language, misspellings meant to dodge filters, and content in any language your members write in. It also reads images and videos, and the text in them. A message that Auto-Moderation has already removed is not checked again by AI.

​

Start in Monitor mode

The Operating mode setting is the most important control on the page.

​

Sensitivity

One slider, expressed as a percentage. The AI scores every message from 0 to 100% for each category — how likely it is that the message breaks it — and a category acts once its score reaches 100 minus the sensitivity: at a sensitivity of 25, from 75%. A balanced starting point is 25–40%. For text, a sensitivity above 50 works like 50. Below a score of 50 the AI judges a message more likely fine than not, so a higher setting would only add false positives. A message that tries to talk its way past moderation — claiming staff approved it, addressing the AI, or excusing itself as a joke or fiction — is always checked at a sensitivity of at least 50, and its log entry says so. Images have no such limit. The AI is usually certain about an image; sensitivity decides the borderline ones, which it scores from 40 to 69%.

​

Categories

Ten categories, each with its own switch, its own set of actions, and its own sensitivity.

​

Age-gated channels

In a Discord age-restricted channel, the Sexual content category is suppressed. Every other category — including sexual content involving minors — still enforces normally. An image or video the AI could not judge is reported to staff there instead of being deleted; see How images are checked below.

​

Actions

Pick any combination per category: Delete message, Warn user, Timeout, Mute, Kick, Ban, Alert. Three combinations are refused:

• Kick and Ban together
• Alert with Ban
• Alert with Kick

Everything else stacks. Delete combines with anything.

​

Custom rules

Write your server’s own rules in plain English. The AI checks them alongside the standard categories: in messages and member names, in what images show and the text in them, and in the speech and on-screen text of videos. Each rule has a name, a description of up to 1,000 characters, a single action, and an optional sensitivity that overrides the global value. As with categories, a sensitivity above 50 works like 50 for text.

​

Escalation

Off by default. When on, repeat offenders receive progressively stronger actions.

​

Exemptions

• Exempt roles — members with any of these are skipped entirely
• Exempt channels — messages here are never scanned
• Exempt users — up to 500, useful for bots

​

Confirmation before kick and ban

On by default. Kick and ban actions are posted for a moderator to approve rather than applied immediately. Confirmations expire after one hour. They are posted to your alert channel, falling back to the log channel, falling back to the channel the message was in.

​

Trust Tiers

A token-saving measure. Members with a long clean record are scanned less often. Four safeguards mean a trusted member is never invisible:

1. Discord invites, @everyone/@here, and 5+ mentions are always scanned — no exceptions
2. A burst of 3 messages in 6 seconds forces a scan
3. Sampling never reaches zero — a trusted member is still scanned periodically
4. Any violation resets trust to nothing instantly

The Trusted Members list shows who currently qualifies. Reduce halves someone’s score; Remove clears it entirely. Both take effect immediately — the save bar is not involved.

​

Channels

• Log channel — every moderation event, with the matched category and confidence
• Alert channel — high-severity events and confirmation prompts

​

When the token balance runs out

AI Auto-Mod is fail-closed: if it cannot confirm there is budget, it does not scan. Notices go to your alert channel, falling back to the log channel. The low-balance warning fires again if the balance recovers and drops a second time.

​

When the AI does not answer

If the AI does not answer for a message’s text — it is down, or too slow — the message gets no AI verdict, costs nothing and is not checked again; a member’s name is checked again at the member’s next update or name change. Standard Auto-Moderation is not affected. Images and videos are checked once more; see below.

​

How images are checked

Up to four images per message are checked, including images in embeds from Discord, Tenor, Giphy and Imgur, and in messages someone forwards. Images past the fourth are not checked; your alert channel is told, at most once an hour.

• Animated images — GIFs and animated WebP — are checked across their frames, sampled from first to last. An animation too long or too large to sample fully, or an animated PNG, is checked on what can be sampled; if nothing is found, your alert channel is told it was not fully checked.
• Videos are checked by what they show and what is said in them; see How videos are checked below.
• SVGs and tiny images — under 10 pixels, or thinner than 1:200 — are skipped.
• Large images — such as photos and full-screen screenshots — are checked at a reduced size that keeps their text readable, which costs fewer tokens.

If the AI refuses to judge an image, or a safety filter blocks it, a second AI model classifies it. If neither can, the image is deleted and staff are alerted; in an age-restricted channel staff are only alerted. The member gets no warning and no infraction for it. A message whose text the AI provider’s safety filter blocks is handled the same way. The AI also reads the text in each image — a screenshot of a chat, a meme’s caption — as part of judging it, and checks some of it again like a message: text in a non-Latin script such as Arabic, Cyrillic or Chinese, or with digits in place of letters; and all of it when the AI is unsure about the image or your server has custom rules, which are checked against that text and against what the image shows. This check can only add a finding, never clear one, and it costs nothing. If it goes unanswered, the image’s own verdict stands. If no AI model answers at all, the image is checked again about 30 minutes later, and your alert channel is told, at most once an hour. An image the AI provider cannot read is not checked again; your alert channel is told the same way.

​

How videos are checked

With Scan images and videos on, videos are checked too: video attachments, and the video that a Tenor or Giphy GIF plays. Videos on sites such as YouTube or Twitch are not downloaded; their preview image is checked like any image. Each video is checked in the background, usually within half a minute of being posted:

• What it shows — up to 12 frames, from the first to the last and including scene changes, are checked like an image.
• What is said and written — the speech in its first 60 seconds is transcribed and checked like a message together with the text on screen, custom rules included. Speech after the first minute is not checked.

Each finding is weighed against your settings, and the most serious one your settings act on is applied, like any other AI Auto-Mod verdict. Frames the AI refuses to judge are handled like such an image: the video is deleted and staff are alerted, or only alerted in an age-restricted channel. Some videos are not checked. Your alert channel is told, at most once an hour:

• Files over 100 MB.
• A video that cannot be downloaded or played. It is not checked again.
• A video the AI could not fully check. If the AI does not answer for part of it, it is checked once more about a minute later. A video whose picture cannot be read still has its speech checked.
• More than 20 videos waiting on one server, or videos from more than 3 messages of one member — the ones beyond that are skipped.

​

What a scan costs

One message produces one AI call, charged the same whatever its length and however many custom rules you have. A message with images produces one call for the text plus one per image, up to four; reading the text in an image is free. An image the first model refuses to judge costs one more call, because a second model then checks it. An image the provider cannot read costs nothing. A video costs up to four calls: up to two for its frames, one to transcribe its speech and one to check what was said and shown on screen, which is free when the video has no speech. A Tenor or Giphy GIF counts as both an image and a video. A video that cannot be downloaded or played costs nothing, and a repost of one already checked usually costs nothing. A video the AI could not fully check the first time is checked once more, which can cost some of those calls again.

​

Username scanning

Display names and usernames are checked when a member joins and when they change their nickname, display name or username. Other member updates, such as a new role, cost at most one check per member every 30 days. Names are checked against your custom rules as well as the categories. A Delete action is meaningless for a name, so it becomes an alert instead.

​

Troubleshooting