Honesty over marketing
Limits of anonymization
No detection system finds everything. We would rather tell you exactly what ours cannot do — with numbers you can reproduce — than sell you a feeling of safety that breaks on the first real case file.
What detection does not do
No completeness guarantee
A hybrid of patterns (with checksums: IBAN, tax ID, health-insurance number …) and AI models for names and free text. A value neither one catches will be missing as a suggestion — which is why the final decision always stays with you; nothing is ever redacted automatically.
Scans need OCR models
Image pages without a text layer run through OCR. Pages with uncertain recognition are explicitly flagged for manual review; if the OCR models are not installed, the export refuses a seemingly clean output (fail-closed) instead of allowing it.
Images inside images
Photos, stamp series, signatures in raster images: the image analysis proposes what it finds — but no system recognizes a handwritten signature in a photo 100 % of the time. Such pages belong in manual review.
Very short values are unsearchable
An age ("45") or a two-digit number cannot be searched reliably without shredding the document. The export lists such cases in the verification report instead of skipping them silently.
Our numbers — both directions
Measured with our benchmark (dm-benchmark) on planted PII in real German case files and on a public German wild corpus. Measured 2026-10-25 with version 0.8.19; the same runs are stored as JSON in the repository.
| Corpus | Documents/pages | Found | Extra suggestions |
|---|---|---|---|
| Own files, digital (planted PII) | 16 / 51 | 158 of 158 (100 %) | 156 |
| Own files, scanned (OCR, planted PII) | 11 / 30 | 122 of 122 (100 %) | 158 |
| German wild corpus (public real texts) | — | ≈ 73 % text recall | — |
The high number of extra suggestions is deliberate: a handful of false positives you dismiss with one click beats one name that slipped through. Nothing is redacted without your decision.
Pseudonymized is not anonymous
- The "Anonymisieren" mode replaces names with per-case labels ([NAME_1]) — that is pseudonymization under Art. 4 No. 5 GDPR, and that is exactly how it is built: with a local mapping that lets you translate AI answers back.
- The moment you send pseudonymized text to an AI provider, the responsibility for that transfer is yours — the tool removes the data from the request, it does not make the transfer lawful.
- A "never redact" dictionary rule only hides pattern matches. What the AI model finds stays visible and flagged — your decision, not the rule's.
What we guarantee
- Everything runs locally on your machine (§ 203 StGB): detection, redaction, verification. No document content ever leaves the device.
- Every export is re-checked after redaction: remaining readable findings block the export or are listed in the report — there is no "looks clean" PDF without verification.
- Every decision is recorded in the locally stored audit trail — traceable who decided what, when.
- Models are verified cryptographically against a signed manifest; tampering fails before anything is processed.
As of 2026-10-25 · version 0.8.19 · measurements reproducible via the benchmark in the repository. Questions about a limit: info@datenmaske.de