← Blog  ·  April 28, 2026  ·  6 min read

OpenAI Privacy Filter — languages supported (2026)

OpenAI Privacy Filter launched with a strong English focus, but many teams process multilingual content: support tickets from European markets, documents from Latin American customers, or logs that mix two languages in the same line. This page covers exactly what language support looks like today, where accuracy degrades, and what you can do about it.

Language support tiers

The underlying model is LLM-based, which means it understands language context rather than matching fixed patterns. This gives it an advantage over regex-based detectors for European languages where personal names and local ID formats appear in natural prose. But it also means accuracy is higher for languages well-represented in LLM training data (primarily English and major European languages).

How multi-language input is handled

You don't need to specify a language parameter — the model auto-detects the input language. If your text mixes languages (code-switching), the model handles it better than regex-based tools because it reads the full sentence context:

import httpx

# Mixed Italian/English — correctly flags both names
resp = httpx.post(
    "https://privacyfilter.run/api/redact",
    json={"text": "Ciao, sono Marco Rossi. My colleague is Sarah Johnson at sarah@acme.com",
          "license_key": "your-key"}
)
print(resp.json()["redacted_text"])
# → "Ciao, sono [PERSON_1]. My colleague is [PERSON_2] at [EMAIL_3]"

Structured PII formats by country

The gap between "experimental" and "strong" support is most visible for government-issued ID numbers, which have country-specific formats:

• US SSN (XXX-XX-XXXX): strong detection
• UK NIN (AA NNNNNN A): moderate; often flagged as OTHER not SSN
• Italian Codice Fiscale (16-char alphanumeric): mostly detected
• German Steuernummer (10–11 digit): partial
• Spanish DNI (8 digits + letter): inconsistent
• Brazilian CPF (XXX.XXX.XXX-XX): partial
• French INSEE (15 digit): sometimes missed in running text

For production processing of non-English ID documents, supplement with a dedicated national-format regex layer. The full entity types reference explains how each type is classified.

Workaround for unsupported languages

If you need reliable PII detection in Japanese, Arabic, Russian, or Hindi, the recommended pattern is to use an LLM-based translation step before redaction — or switch to a self-hosted tool with proper multilingual support like Microsoft Presidio with a matching spaCy language model.

import httpx

def redact_non_english(text: str, license_key: str) -> dict:
    # Step 1: translate to English via an LLM (Claude, GPT-4o, etc.)
    # Step 2: redact the English version
    # Step 3: map entity offsets back to original text (optional)
    en_text = translate_to_english(text)  # your translation function
    return httpx.post(
        "https://privacyfilter.run/api/redact",
        json={"text": en_text, "license_key": license_key}
    ).json()

FAQ