Datenmaske
AI TRANSPARENCY

Which AI sees
your document?

We disclose which models work inside Datenmaske, where they run and what they were trained on. Inference exclusively in the EU, no training on your documents , the final redaction decision always made by the user.

EU AI REGULATION

Datenmaske is a system with a minimal risk class.

The EU AI Regulation (AI Act) classifies AI systems by risk. Datenmaske does not fall under the Annex III high-risk categories (no biometrics, no law enforcement, none of the critical use areas listed there). PII detection in text documents is tool-supportive and always subject to human approval.

Classification: minimal risk class (comparable to spam filters). We do not claim an “AI Act certificate” — no such thing exists for minimal risk classes. The correct statement is: minimal risk + voluntary transparency disclosures.

Date
AI Act milestone
Status
02/2025
Prohibited practices (Art. 5) + AI competence duty (Art. 4) apply
in force
08/2025
GPAI model obligations (Chapter V) apply
in force
08/2026
High-risk obligations + Art. 50 transparency apply
upcoming

MODEL CARDS

The models in detail

spaCy NER (de_core_news_lg)

Purpose
Detection of German proper names (people, places, organisations) in the text layer
Origin
Open-source model (Explosion / spaCy), pre-trained on German corpora
Inference location
Self-hosted on spacy.mokka-dev.de (EU)
Training on customer documents
No — never.
Updates
Version-pinned; no automatic re-training on customer data

GLiNER (self-hosted)

Purpose
Zero-shot Named Entity Recognition to extend the spaCy detection
Origin
Open-source model
Inference location
Self-hosted in the EU
Training on customer documents
No — never.
Updates
Version-pinned; no training on customer data

urchade/gliner_multi_pii-v1 (PII-NER)

Purpose
PII-specific zero-shot NER to extend the detection of special data categories (health, financial data, identity documents)
Origin
Open-source model (urchade/gliner_multi_pii-v1, mDeBERTa), available on HuggingFace / onnx-community
Inference location
Self-hosted in the EU (onnx-community ONNX runtime)
Training on customer documents
No — never.
Updates
Version-pinned; no training on customer data

OpenMed (medical NER)

Purpose
Medical Named Entity Recognition for doctor's letters, diagnoses and other health data (Art. 9 GDPR)
Origin
Open-source model (HuggingFace), medical NER corpus
Inference location
Self-hosted in the EU
Training on customer documents
No — never.
Updates
Version-pinned; no training on customer data

Azure Document Intelligence (OCR)

Purpose
Text recognition exclusively for scanned pages (PDF images)
Origin
Microsoft Ireland Operations Ltd.
Inference location
Region Germany (Germany West Central); EU Data Boundary
Training on customer documents
No — never.
Updates
Microsoft-side; no use for model training (contractual)

Further components of the detection pipeline (rule-based patterns/gazetteers, string matchers) are not learning models and are documented in the pipeline overview at /forschung. The model cards above describe the production NER and OCR stack of Datenmaske Web; the Desktop edition (§ 203) uses a partly different, locally bundled model selection.

ASSURANCES

What we assure — and what we do not.

  • EU inference. All model inference runs in the EU — spaCy/GLiNER self-hosted (Germany), Azure OCR in the Germany region (Germany West Central). No transfer to third countries takes place (see DPA § 2 Abs. 3).
  • No training on customer documents. We never use your documents to train, fine-tune or improve our models. Models are version-pinned and maintained independently of customer data.
  • Human-in-the-Loop, fail-closed. Every redaction requires approval by the user. Export is blocked until verification (no detected PII string remaining in the text layer) has passed.
  • Audit trail. Every detection and redaction is traceably documented in the redaction log (with SHA-256 hashes) — Art. 5 Abs. 2 GDPR accountability.
  • No LLM chat feature at present. Datenmaske currently does not deploy any generative Large Language Model that directly “interacts” with end users (Art. 50 Abs. 1 AI Act). The AI runs as background processing with human approval. Should an LLM feature be added, the Art. 50 transparency obligation will be fulfilled in advance.
  • No guarantee of detection completeness. AI-based suggestions must be reviewed by the user. Datenmaske supports manual review; it does not replace it.
DATENMASKE

Transparency you can verify.

Upload a PDF and check detection, approval and redaction log yourself — free, no registration.