AI moderation on Discord: what it can and cannot do
Image models, language models and Discord's built-in filters, with the mistakes each one makes and the part your moderators keep.
By SfwBot Team7 min read

AI moderation on Discord works well for narrow jobs: an image model can rate every picture for nudity and delete the explicit ones at any hour, and a language model can flag hostile messages that contain no banned word. It cannot weigh context the way your team does, so it makes mistakes in both directions, and a person still has to decide appeals and anything that depends on who said what to whom.
The kinds of automatic moderation on Discord
Three different things get called AI moderation:
- Discord's own tools. AutoMod blocks messages that match your word lists, Discord's lists of commonly flagged words, or content that resembles widely reported spam. The Explicit image filter under Server Settings, Safety Setup checks images for explicit content. Both are free, built into every server and stop a message before other members see it. Switch them on before you add any bot.
- Image classifiers. A bot downloads each image and a trained model rates it. SfwBot works this way.
- Language models. A bot passes message text to a model that labels it, for example as harassment or a threat. Some bots run their own model, others send the text to an AI company's service.
What an image model does, with SfwBot as the example
SfwBot's nudity detection rates each image, GIF and video on four levels: Everyday, Revealing, Implied nudity and Explicit. You pick the level where blocking starts, and the bot deletes everything at that level and above. A new server blocks Explicit only. For a GIF or video the bot rates a sample of frames.
The model only rates. Your level, your exempt channels and roles, and your allow and block lists decide what happens to the message. How AI NSFW detection works follows one image through every step.
SfwBot detects nudity only. It does not look for violence, gore or other harmful images, so those stay with your moderators.
Text moderation with language models
A language model can catch an insult that contains no banned word or is spelled to get around a filter, which a word list cannot.
Its weak spots come from what it cannot see. Friends insulting each other as a joke, a quote posted to report someone, sarcasm or slang the model knows badly can all be flagged, and a real threat in ordinary words can pass. A text model also reads every message, so far more of your members' writing leaves Discord than with an image check.
SfwBot does no AI text moderation. Its invite filter, spam filter and link protection follow rules and lists you control, use no credits and keep only the outcome of each check.
Wrong calls in both directions

A false positive removes something harmless. A false negative lets through something the model should have caught. Every model makes both.
With images, the usual false positives are close-ups of skin-coloured surfaces such as hands or food, classical paintings and borderline outfits. The usual misses are small or cropped images, and explicit frames in a long video that fall between the frames the bot samples.
The level you choose moves the balance. Blocking from Revealing also removes swimwear photos, harmless beach pictures included. To see each case before anything is deleted, switch off Delete blocked content: the bot then only reports to your log channel and a moderator decides. When an image slips through, add it to your block list so its fingerprint is caught the next time.
Context a model cannot see

An image model looks at the picture alone. It does not read the caption, know the channel's topic or know who posted.
- Art. Drawings, paintings and sculpture are rated by what they show. On Plus and Pro, Illustrations and art gives artwork its own level or allows it entirely.
- Medical and educational images. An anatomy diagram or a photo for a health question is rated like any other picture. Hold those discussions in a channel you add to Exempt channels, or on Pro give that channel its own level.
- Memes. A crude joke with no nudity in it is Everyday to the model. Whether the joke breaks your rules is for your moderators to judge.
Where your members' posts go
With any AI moderation, a copy of what members post leaves Discord to be checked. Find out who runs the model, whether an outside AI company receives the content, and what is stored for how long. What to check before you add a bot lists the questions.
SfwBot downloads each image to scan it and discards it right after. It keeps a fingerprint of the image and the link to the media on Discord for 30 days; fingerprints on your own lists stay until you remove them. Its servers are hosted in Germany, and the list of processors on its security and privacy page names no outside AI company. Staff see an image only when a decision is reported for review, by opening that one image through the stored link.
Cost and speed
Every AI check uses computing time that the operator pays for. With SfwBot, each AI scan uses credits, and the free plan includes credits every month. Matches on your allow and block lists cost nothing, and the same media posted again within a day is not scanned twice. When the credits run out, scanning pauses and your block list keeps working.
A bot learns about a message only after Discord has posted it. While the bot downloads and rates an image, the image is in the channel and members who are looking can see it. Discord's explicit image filter and AutoMod act before the message appears, so keep them on next to any bot.
Why moderators still decide appeals
A model gives a rating without knowing your rules, the member's history or what they meant. Before you rely on a bot, check how a member and a moderator can undo a wrong call.
In SfwBot, each removal goes to your log channel. Allow in this server puts the image on your allow list so it is not removed again, and Report as wrong sends it to the SfwBot team for review. Members can report their own removals through /standing. A removal by the AI scan costs the member no trust unless you switch that on in the trust settings.
Bans, arguments between members and every case someone disputes stay with your team. The nudity detection guide covers each setting and button.
Frequently asked questions
Can AI moderate a Discord server on its own?
No. It handles narrow, repetitive checks such as removing explicit images at any hour. Rules that depend on context, disputes between members and appeals need a person.
Does SfwBot use AI to read text messages?
No. SfwBot's AI rates images, GIFs and videos for nudity. Invites, spam and scam links are handled by filters that follow rules and lists, and they use no credits.
Is there free AI moderation for Discord servers?
Discord's AutoMod and explicit image filter are free and built into every server. Bots that run their own models have costs to cover, and SfwBot's free plan includes scan credits every month.


