Implementation guide
OWASP AISVS v1.01
Input Validation · Content & Policy Screening
Check screening works in other languages
Verify that prompt content classification is evaluated for unsupported languages.
OWASP AISVS v1.01, requirement 2.2.2
In plain English
Find out how well your content screening works in languages you do not officially support. Classifiers are usually strongest in English, and switching language is one of the oldest ways round a filter.
What good looks like
- Screening has been tested in a range of languages beyond the supported ones, including some with few speakers online.
- There is a written rule for input in a language the classifier handles poorly: block it, translate it first or apply stricter handling.
- The tests are rerun when the classifier or its version changes.
How teams usually do it
- Language detection in front of the classifier (for example fastText's language identification model or a cloud equivalent), with a policy per language.
- A multilingual test set that covers each harmful-content category, including mixed-language and transliterated prompts.
- Translating unsupported-language input to a supported language and screening both versions.
Suggested evidence
- Test results broken down by language.
- The written policy for unsupported languages.
- Log samples showing the policy being applied.
Common mistakes
- Testing in English and assuming the classifier is as good elsewhere.
- Ignoring code-switching, where a prompt mixes two languages mid-sentence.
- Forgetting transliteration, such as Hindi or Arabic written in Latin letters, which language detectors often misread.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.