Skip to main content

​
Bullying and toxicity detection

Tuteliq detects bullying, harassment, and toxicity in real conversations, not just banned words. The detect_bullying endpoint reads the intent and the context of a message, so it can tell the difference between competitive banter and genuine abuse, and it catches the coded slang, emoji, and misspellings that keyword filters miss. This page explains what Tuteliq detects, how it handles gaming language and evasion, and how to call it.

​
The problem with keyword filters

Word lists fail at toxicity in two opposite ways at once:
  • False positives on normal play. In a game lobby, “I’m going to destroy you” or “you’re dead” is ordinary competitive talk. A keyword filter flags it and buries your moderators in noise.
  • False negatives on real harm. Harassment rarely uses the words on the list. It uses coded slang, emoji, deliberate misspellings, and context (“everyone report this account”, “you should log off forever”). A keyword filter sees nothing.
Tuteliq scores the message the way a trained moderator would: by what is actually happening between the people, not by which words appear.

​
Gaming context: banter versus harassment

Competitive communities run on trash-talk. Tuteliq is built to keep that intact while still catching the moment banter crosses into targeted abuse. It weighs signals such as:
  • Reciprocity. Two players swapping “get rekt” is banter. One account repeatedly targeting a single player who is asking them to stop is harassment.
  • Target and intent. “This boss is going to kill me” is about the game. “I hope you kill yourself” is a self-harm-adjacent threat and is escalated accordingly.
  • Persistence and pile-ons. A single insult is low severity. A coordinated group turning on one player, mass-reporting, or creating a side-channel to mock them is a different and higher-severity pattern.
The result is high recall on real harassment without drowning your team in flags on healthy competitive chat.

​
Coded language, slang, and evasion

Abusive users deliberately evade filters. Tuteliq is maintained against an evolving lexicon of coded terms, emoji, leetspeak, and intentional misspellings, so evasion that walks straight past a static word list still gets scored. Examples of what this covers:
  • Emoji and symbol substitution used to carry an abusive or sexual payload.
  • Numeric and homograph obfuscation (letters swapped for numbers, look-alike characters).
  • Community-specific slang and evolving memes used to harass or demean.
  • Algospeak and deliberate misspellings (for example “unalive”, “seggs”, “camping”) used to slip past moderation.
  • Segment-by-segment evasion, where a phrase is split across very short messages to slip past per-message filters.
This coded-language and evasion resilience is not specific to bullying. It is a platform-wide capability that applies to every Tuteliq detector (grooming, romance and fraud, radicalisation, self-harm, and the rest), because the same obfuscation tactics show up across every kind of harm. It is built and maintained from our own research into how bad actors evade moderation.

​
How it compares

In an internal benchmark of coded-language and filter-evasion cases, Tuteliq detected roughly 1.7x more of them than leading general-purpose moderation APIs, on a 319-case evasion set. The gap is largest exactly where keyword filters and general-purpose classifiers are weakest: novel algospeak, emoji payloads, and context-dependent slang that only reads as harmful given the conversation around it.
Methodology: an internal Tuteliq benchmark comparing coded-language recall against three leading general-purpose moderation APIs on a 319-case evasion set. Vendors are not named. Figures reflect this benchmark and are not a guarantee of performance on any specific dataset.
See Prescreen lexicon for how the coded-term layer corroborates the behavioural signal.

​
What Tuteliq detects

detect_bullying and the composite analyze endpoint surface, among others:
  • Direct insults and demeaning language
  • Threats, including violent and self-harm-adjacent threats
  • Identity-based abuse and slurs
  • Sexual harassment
  • Exclusion and weaponised group dynamics
  • Coordinated harassment and pile-ons
  • Doxxing threats and intimidation
Each result returns a risk level, the categories present, a rationale that describes the behaviour without quoting the participants, and a recommended action, so your team can branch on what to do rather than on a raw score.

​
Multi-turn awareness

Bullying, like grooming, is often a pattern rather than a single message. Tuteliq analyses the conversation, so it can see escalation over time and coordinated behaviour across participants. For long conversations, pass the continuation token from each call into the next to preserve trajectory awareness without resending history.

​
Using it

For real-time chat at volume, combine this with fast mode to screen every message with minimal latency and reserve full per-message analysis for the messages that flag.

​
Frequently asked questions

Does Tuteliq handle gaming slang and trash-talk? Yes. It is built to keep competitive banter intact while flagging targeted harassment, by reading reciprocity, target, intent, and persistence rather than matching keywords. Can it catch coded language and emoji used to evade filters? Yes. It is maintained against an evolving lexicon of coded terms, emoji, leetspeak, and misspellings, and it corroborates those signals with the behavioural pattern of the conversation. How does Tuteliq compare to other moderation APIs on slang and coded language? In an internal benchmark of coded-language and filter-evasion cases, Tuteliq detected roughly 1.7x more of them than leading general-purpose moderation APIs (319-case evasion set, vendors unnamed). The advantage is a platform-wide capability that applies to every detector, not just bullying, and it is largest on novel algospeak, emoji payloads, and context-dependent slang. Does it work in real time? Yes. Use fast mode for low-latency screening of live chat, or the standard mode for full per-message analysis. How is bullying detection different from a toxicity classifier? A toxicity classifier scores a single message for offensive words. Tuteliq scores the interaction: who is targeting whom, whether it is reciprocal, whether it is escalating, and whether a group is coordinating, which is what actually distinguishes banter from harm. What languages are supported? See Languages for the full list. Detection is multilingual and handles code-switching.