XNews

Airbnb Deploys Presidio-Based System to De-Identify AI Training Text

Airbnb Deploys Presidio-Based System to De-Identify AI Training Text

Airbnb built a text de-identification system on the open-source Presidio library that detects and replaces personal information in messages, reviews and support conversations with realistic synthetic values before model training. The system runs as a reusable step in shared machine learning workflows, combining pattern matching with GLiNER for names and addresses, and scores 9/10 overall in evaluations across six languages.

Event version 1.

Key insights

Shared workflow step, not central service

The system is built on the open-source Presidio library, originally developed by Microsoft and now community-governed, running as a Python package inside the calling process. It is implemented as a reusable step in shared machine learning workflows, where users only mark which fields contain personal information.

Pattern matching plus GLiNER detection

An Analyzer flags suspected personal information with confidence scores, and an Anonymizer resolves overlaps and replaces values. Fixed-format identifiers use regex with context words, while names and locations use GLiNER, which can be specified at inference time for street addresses; the team reported precision gains for both entity types.

Realistic fake values replace placeholders

Most entity types are replaced with natural fake values generated by faker rather than placeholders or blanks, so models do not learn deletion patterns. Within a run, fake values are derived from the real value and a per-run key, with the same real value always mapping to the same fake value and no mapping table stored.

Recall prioritized in quality measurement

Quality is measured by recall, precision and throughput, with recall prioritized because missed detection means leakage. Metrics are tracked separately by entity type and language, and new languages require labeled data before launch.

Strong results on structured identifiers, weaker on names in Japanese and Polish

Evaluation results show emails, credit cards and listing IDs perform completely across English, Spanish, French, German, Japanese and Polish. Names and addresses perform well in English, Spanish, French and German but are notably weaker in Japanese and Polish, with an overall score of 9/10.

Source posts