AI detectors: dates, addresses, amounts and QR in PDF

Detecting national IDs and email is necessary but not enough. In resolutions, payrolls, clinical records and judgments, re-identification often comes through the side door: a date of birth, a full home address, an amount only one person receives, or a QR/stamp linking to an external verifier.

Anonimatum includes AI detectors aimed at these cases — dates, addresses, amounts, QR/stamps — running with proprietary models in the EU. This guide explains *when* to enable them, how to avoid over-redaction, and how to combine them with patterns and human review.

Dates: not all are equal

A document date («Madrid, 12 March 2026») may be harmless. A birth date, surgery date or police summons date, combined with initials or neighbourhood, may identify someone. Practical rule:

• Keep administrative dates needed to understand the act (unless policy says otherwise).
• Prioritize redacting vital / clinical / sanctioning dates when publishing.
• Review deadline tables: risk is sometimes in the column, not the heading.

Addresses: the silent identifier

In transparency and notices, the home address is often «forgotten» because it looks like context. A portal with the name redacted but the address intact still exposes the person. Partial addresses (street + number without city) also matter in small municipalities.

• Decide whether you publish with address omitted, partially masked or replaced.
• Watch letterheads, footers and map annexes.
• Combine the address detector with local place-name word sets if your territory requires it.

Amounts: when money identifies

An amount is not PII in the abstract. It is when, in context, it can only refer to one person (singular compensation, fine, payroll of a unique role, invoice of a named freelancer on another page). In due diligence, amounts also reveal strategy; document whether your goal is privacy, trade secrecy or both.

QR and stamps: the link outside the PDF

A QR code may open verification with a name, document hash or appointment portal. A digital stamp or watermark may embed identifiers. Treating only OCR text leaves the graphic code intact. That is why a dedicated detector is needed and, often, visual zones over the symbol.

• Inventory which QR codes are signature, payment, tracking or citizen verification.
• Decide whether to remove the code, replace it with a marker or hide the region.
• Verify the output PDF cannot be scanned for the original QR (test with a phone).

How to avoid over-redaction

Enabling every detector «just in case» yields useless documents and internal pushback. Better approach:

• Start from a policy by document type (public resolution ≠ internal payroll ≠ training material).
• Use preview to whitelist what must remain (e.g. procedure dates).
• Complement AI with custom patterns for local formats — not the other way around.
• Measure false positives on a sample of 20–50 real documents before mass rollout.

Combining with batches and signatures

In remittances, detectors apply uniformly: document the policy in the batch queue. If the PDF is signed, remember anonymization may affect the signature; Anonimatum provides warnings and, when the process allows, verification/re-sign flows.

Summary for DPOs / owners

Date, address, amount and QR detectors are not marketing features: they are controls against indirect re-identification. Enable them with policy, measure them on samples and require preview. In Anonimatum they run with local AI in the EU, aligned with the rest of the document stack.