Automatic comment moderation: how the AI decides
Automatic does not mean blind. Here is what happens in the seconds after a comment is posted: what the AI reads, how it weighs context and irony, where the thresholds sit, and which cases it deliberately leaves to a human.
Trusted by
The problem
Word lists judge the letter of a comment, not its meaning. They hide “our defence was murdered tonight” and let “go back where you came from” through. And a filter that hides comments without saying why cannot be defended to anyone.
How OpenModeration handles it
Our AI reads the comment with what surrounds it: the article or post, the message it replies to, its language. It scores each category of your rules, then compares each score with two thresholds: below the first it publishes, above the second it hides, in between a person decides.
The result
Clear cases are settled within seconds, day and night. Uncertain ones go to our moderators instead of being guessed. Every decision, automatic or human, comes with a sentence that explains it.
What the AI reads before deciding
Four signals, read together. None of them is a list of forbidden words.
The page
The same sentence does not mean the same thing under a match report and under an article about a crime. The AI reads the topic before the comment.
The conversation
A reply is read with the message it answers. Sarcasm aimed at an argument is not treated like an attack on a person.
Irony and coded language
Disguised spellings, emojis used as insults, dog whistles: the AI looks at what is meant, not only at what is written.
Your rules
Each category is judged against the rules set with you in plain language: what you accept, what you refuse, where you want caution.
Six comments, six decisions
Illustrative examples: the same AI reaches different outcomes depending on the context and on your rules.
Under a football match report
“Our defence was murdered tonight, what a disaster.”
Published
Violent word, sporting meaning. Criticism of a team's performance, no person targeted.
Under an article about a new minister
“Great, another genius in charge. Can't wait.”
Published
Ironic, but aimed at a public figure's role. Political criticism is allowed under your rules.
Reply to another reader
“People like her should be sent back where they came from.”
Hidden
Exclusion based on origin, aimed at a person. Breaks the rule on hateful content.
Under an Instagram post
“Earn €500 a day from home, DM me 💰”
Hidden
Commercial spam asking for contact. Breaks the rule on scams and advertising.
Under a local news story
“We all know what goes on in that neighbourhood at night…”
Sent to our team
Possible insinuation about a group, with no explicit target. The score falls between the two thresholds: a moderator decides.
Under a city council article
“The mayor lied about the budget and should resign.”
Published
A harsh accusation against an elected official, in a political debate. Neither an insult nor a threat.
Thresholds: where automatic stops
For each category of your rules, two thresholds split comments into three outcomes. We set them with you and adjust them after the test on your comments.
| Category | Published automatically | Sent to our team | Hidden automatically |
|---|---|---|---|
| Insults | Rough tone, no target | Insult that may be a quote or a joke | Insult aimed at a person |
| Hate | Debate on a sensitive topic | Ambiguous generalisation about a group | Attack based on origin, religion, gender… |
| Threats | Figurative violence (“I'll kill this exam”) | Unclear intent | Explicit threat against someone |
| Spam and scams | Link relevant to the discussion | Repeated promotional message | Fraudulent offer, phishing |
| Off-topic | Side remark that stays civil | Long unrelated message | Flooding, copy-paste campaigns |
Stricter or looser? You choose per category: a busy news site and a brand page rarely set the same thresholds.
Why teams choose OpenModeration
It does not guess
When the AI is unsure, the comment goes to our moderators, who decide according to your rules.
Reports and appeals stay human
User reports and contested decisions are reviewed by a person on our team, as the DSA requires.
You keep the last word
You see every automatic decision in real time with its reason, and you can override any of them.
Frequently asked questions
How does automatic comment moderation work?
Each comment is read with its context, scored against each category of your rules, then compared with two thresholds: publish, hide, or send to a person. The whole process takes a few seconds.
Does the AI understand irony?
Much better than a word list, because it reads the comment together with the page and the conversation. When irony makes a comment truly ambiguous, its score falls between the thresholds and our team decides.
What happens when the AI gets it wrong?
You can reverse any decision from the console, and authors can appeal. Recurring errors are used to adjust your rules and thresholds with you.
In which languages does it work?
Comments are analysed in their own language, including the languages of your audiences, without a translation step that would lose nuance.
Can I check the decisions before switching it on?
Yes. A free sandbox test on your real comments shows every decision and its reason before anything goes live.
Watch the AI decide on your own comments.
Book a call: we run your recent comments through the AI and walk you through each decision and its reason.
No credit card required · DSA-compliant · GDPR-compliant
Questions? contact@openmoderation.com