Engine changelog

Seventeen languages, each with its own list

Every language has a hand-curated file: its terms, the phrases that turn a word into a threat or an insult, and the words that show an innocent sense.

Reads the disguises people type

Leetspeak, look-alike letters and signs from any alphabet ("ƒυ¢к", "₣₳₲₲Ø₮"), dots and spaces between letters, a letter masked or left out ("f ck"), stretched letters, words written backwards, letter-shaped emoji, Arabizi digits and Korean typed in English. Each disguise read is named.

Emoji that say the word

A picture in a sentence reads as its word ("piece of 💩", "I'll 🔪 you", "you're a 🤡"), the middle finger as what it means, and pairs like 🍆💦 and 🐒🍌 as what neither picture says alone. One aubergine is still dinner.

Threats and dehumanizing statements, not only words

Direct threats against the reader or their family, wishes of violence and claims that a group are not people are caught as phrases, in every language.

A quoted slur counts as a slur

A slur or a swear word quoted, reported or named as a word ("he called me a …") counts as what it is. A slur someone claims for their own group goes to the model.

Usernames read as names

A name run together is split in its own language (jemoederaandepoeder reads as Dutch), a letter in a shape its spelling never takes is read for what it stands for (HitIer, with a capital I), and the model is told which words the rules found in the name.

Lessons and fiction read as what they are

In the educational context a word quoted or named is discussed, and in the creative context a quotation is a character's speech, in every language's quotation marks. A word used on someone still counts, and a slur counts in fiction.

Game chat reads as a game

In the gaming context a taunt about play ("gg ez, uninstall") is not acted on, under every policy but kids. A slur, a threat, a personal attack and "kys" still are.

Of 14 game-chat lines read by the model on 2026-10-08, taunts were rated 0.8 to 1.6 for severity and personal attacks 2.5 to 3.4.

Innocent words stay innocent

Place names, surnames and ordinary words with a profane string inside pass: Scunthorpe, Dickinson, cockatoo, Püssi in Estonia. A word used in its literal or medical sense, negated or reported is read that way.

Quoting abuse to push back is not abuse

Counter-speech, a report of what someone said ("he yelled … at me") and a claim held up to argue with ("people who say that … should be ashamed") go to the model instead of being flagged outright.

Joking threats in games

"I'll kill you next round", "you're dead" said of a match, or a threat said laughing goes to the model instead of being flagged outright. A threat to a family, a rape threat or an incitement is always flagged.

Rudeness, as its own category

Rude, dismissive or harsh text that stops short of an insult, like "shut up" or "what nonsense". A negative opinion put plainly ("our politicians are incompetent") is not rude. Only the model reads it.

Sexual content involving minors, and self-harm

sexual_minors is blocked by every preset. self_harm covers the writer's own self-harm or suicidal thoughts, plans or acts, and every preset holds it for review. Only the model reads either.

A model reads what the rules can't settle

In balanced mode the model reads every text the rules do not flag outright, clean-looking text included, since no word list sees abuse written in ordinary words. It is asked what the writer does, and the policy decides.

Long text stays fast

Text written without spaces, such as a long Chinese, Japanese or Korean passage, takes time in step with its length.