Autopilot
Autopilot looks at the images, GIFs, videos and links people post in your server and removes the harmful ones: nudity, gore, hate symbols, scams and the like. How bad a post has to be, and how sure Autopilot has to be, is up to you, category by category. What it keeps can be logged for a moderator to look at. Automod handles text (banned words, spam, invites); Autopilot handles what people see.
Overview
Autopilot is off until you turn it on. Once it is on, every image, GIF, video and link in a message is checked by an AI model. It answers with what the picture shows and anything in it that breaks a category you've switched on, rated two ways:
- How bad: severe is the worst (sexual acts, real deaths, calls for violence, praise for terror attacks, login scams); serious is plainly harmful (nudity, slurs and racist memes, hate symbols used as a joke); mild is offensive but low harm (stereotypes, edgy jokes, inflammatory claims about a group).
- How sure: sure, or unsure when it's ambiguous, partly hidden or could be read either way.
Before anything is removed, a second, separate check has to agree. A photo of a headphone box doesn't get deleted because one check had a bad moment.
Off, Watch only and On
- Off: nothing is checked.
- Watch only: everything is checked, and what it would have removed is posted to your log channel as Would remove. Nothing is deleted and nobody is messaged.
- On: posts your levels say to remove are removed (and the author is told, if you want).
Start with Watch only. After a few days of Would remove posts you'll know whether its calls match your server's, and you can switch to On.
Categories and levels
There are six groups, each with its own categories. Every category has its own level, and setting a group sets all of its categories at once.
- Sexual content: nudity and sexual acts; suggestive; sexualised minors.
- Violence and gore: gore and injury; violent acts; animal cruelty.
- Hate and extremism: hate symbols; extremist propaganda; hateful text and slurs; hateful caricatures.
- Harassment: targeted abuse of a real person; personal information (doxxing).
- Self-harm: self-harm or suicide shown or encouraged.
- Scams: fake Nitro and giveaways, crypto scams, QR codes asking people to log in or pay.
Hate symbols are judged by how they're used. A swastika shown approvingly or as a joke is removed; a clearly historical or educational image (a WWII history book, a documentary still) is allowed. Hateful text needs a slur or dehumanising words to count as serious: an inflammatory claim or a conspiracy post about a group without one is mild.
Sexualised minors covers any sexual or sexualised depiction of a minor or a childlike character, drawn or real. Discord bans it, so it's always rated severe and can't be set below Standard, and no copy of it is ever kept or shown: its log posts and dashboard entries say what was found, without the media.
Each category is at one of four levels. A lower level is more forgiving twice over: it acts only on worse posts, and only when it's sure.
| Level | Removed | Kept and logged for a moderator |
|---|---|---|
| Off | nothing | nothing |
| Lenient | severe, when sure | nothing |
| Standard | serious or severe, when sure | anything unsure, and mild when sure |
| Strict | anything sure, and serious or severe when unsure | mild, when unsure |
A racist meme is serious: Lenient leaves it up, Standard and Strict remove it. A post calling for violence against a group is severe and goes at every level. Kept posts show up in the log channel as Borderline · kept.
Presets
- Recommended: Standard for nudity, sexualised minors, gore, violent acts, hate, self-harm and scams. Suggestive images, animal cruelty and harassment are off.
- Strict: everything on at Standard, with nudity, suggestive images, sexualised minors and hate at Strict.
- Relaxed: the main categories at Lenient (sexualised minors stays at Standard).
Picking a preset sets every category. Change one category afterwards and the preset reads Custom.
Setting it up
Everything is on one screen, in Discord or on the dashboard:
/autopilot # mode, preset, log channel, bypass roles, what to check- Mode: Off, Watch only or On.
- Preset: Recommended, Strict or Relaxed.
- Log channel: pick one, or let kesarAI create a private one only moderators can see. Removals are posted there with a copy of the media.
- Bypass roles: members with any of these roles aren't checked at all.
- What to check: images, videos, links.
- What to do: delete the post, and message the author.
/quicksetup can do this for you: it creates a log channel (or uses one you pick) and turns on Recommended in Watch only. The /autopilot command needs Discord Administrator; on the dashboard, server admins can change it.
Channels
Every channel uses the server's levels. For any channel you can make an exception: skip it entirely (nothing there is checked or counted against the budget), or give some categories their own level there, like turning off Suggestive in an art channel.
Checking its work
Every removal is posted to the log channel with the reason, what the image looked like, and a copy of it. The post has buttons for moderators:
- Mark as mistake on a removal: the same image is ignored from then on, and the mistake is counted.
- Remove or Looks fine on a borderline post or a Would remove: Remove deletes the original if it's still there; Looks fine lets the same image through from then on.
To test an image, right-click the message and choose Apps → Check with Autopilot. Only you see the answer: what it found, how bad and how sure, and what your settings would do. It never acts.
Activity on the dashboard
The dashboard's Autopilot page has three tabs:
- Overview: what was checked, removed and kept as borderline over the last 24 hours, 7 days or 30 days, by hour or day and by category, and the most recently flagged items.
- Members: the members whose posts were removed most, with their user ID. Tap one to see what of theirs was removed, and why.
- Settings: everything above, plus channels and limits.
Removed media is shown from the copies in your log channel, through Discord's own links, and blurred until you tap it. The website doesn't keep any of it. When there's no copy (a file too big to re-post, a category that is never copied, or a log post that was deleted), the item says so and links to the log post instead.
What members are told
With Message the author on, they get one DM after the post is gone, for example:
Your image in Saffron House was removed by Autopilot: it looked like a hate symbol. If that's a mistake, a moderator can review it.
Watch only never sends one.
Budget and limits
- Daily budget: the most checks in a day, to cap the cost. When it runs out, nothing more is checked until the next day, and the dashboard says so.
- Smallest image and smallest side: stickers, emoji and avatars below these are skipped.
- Banned words in images: text in an image is compared with Automod's banned words, and a match is always removed (you listed those words yourself).
- Post every scan: also posts clean checks to the log channel. Off by default; only removals and borderline items are posted.
What isn't checked
- Direct messages to the bot.
- Members with a bypass role.
- Channels you've set to skip.
- Images under the size limits.
- Posts by bots, unless Check bots' uploads is on. kesarAI never checks its own posts, except what a member sends through
/say: the file, image link and link previews are checked as that member's, and in On mode a post that would be removed isn't sent. - Links, when the link check isn't available. The dashboard shows that while it lasts, and those links aren't counted as clean.
