Skip to main content

​
Benchmarks and accuracy

Tuteliq is built to catch the harm that general-purpose moderation misses: coded slang, emoji, algospeak, and deliberate filter evasion, read in the context of the conversation rather than matched against a word list. This page summarises how it performs, how that is measured, and how it compares.

​
Headline result

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 messages around it.

​
What is measured

The evasion benchmark is a set of messages that carry a harmful payload through obfuscation rather than plain words: coded acronyms, emoji substitution, leetspeak, homoglyphs, deliberate misspellings, and algospeak (for example “unalive”, “seggs”, “camping”). It also includes benign look-alikes that use the same surface vocabulary, so a detector cannot score well simply by flagging everything.
  • Positive cases: messages where an obfuscated harmful meaning is present.
  • Negative controls: benign messages that share vocabulary or symbols with the positives.
  • Metric: recall on the positive cases (what fraction of real evasion is caught), reported alongside behaviour on the negative controls.

​
Methodology

  • Comparison is against three leading general-purpose moderation APIs, averaged. Vendors are not named.
  • Figures reflect this internal benchmark and are not a guarantee of performance on any specific dataset or deployment.
  • The advantage is a platform-wide property of how Tuteliq reads content, so it applies across every detector (grooming, romance and fraud, radicalisation, self-harm, bullying), not only to toxicity.

​
Why the gap exists

General-purpose moderation scores a single message against a fixed vocabulary. Tuteliq scores the interaction: who is targeting whom, whether it is reciprocal, whether it is escalating, and whether the wording performs a harmful function even when no banned word appears. That is why it separates gaming trash-talk from targeted harassment, and why obfuscation that walks straight past a word list still gets scored. See Bullying and toxicity detection and the Prescreen lexicon for how the coded-term layer corroborates the behavioural signal.

​
Multilingual coverage

Detection runs in 27 languages (English stable, all 24 EU official languages plus Ukrainian, Norwegian, and Turkish in beta), with code-switching support and language-specific coverage of slang and evasion. See Languages for the full list.

​
Frequently asked questions

How accurate is Tuteliq at detecting coded slang and filter evasion? On a 319-case evasion benchmark, Tuteliq detected roughly 1.7x more coded-language and filter-evasion cases than leading general-purpose moderation APIs, while still separating benign look-alikes that share the same vocabulary. How is the benchmark measured? It uses obfuscated harmful messages (coded acronyms, emoji, leetspeak, algospeak, misspellings) plus benign look-alike controls, and reports recall on the positives alongside behaviour on the controls. Comparison is against three leading general-purpose moderation APIs, averaged and unnamed. Does the coded-language advantage apply to more than bullying? Yes. It is a platform-wide capability that applies to every detector, because the same obfuscation tactics appear across every kind of harm. How many languages does Tuteliq support? 27: English is stable and 26 more are in beta, including all 24 EU official languages plus Ukrainian, Norwegian, and Turkish.