Airbnb details how it strips personal data from text before training machine learning models
In an engineering blog post, Airbnb says it built its text de-identification system on the open-source Presidio library, swapping in GLiNER for name and address detection and replacing personal data with realistic fake values. The claims are Airbnb's own and have not been independently verified.
Airbnb says it now uses transaction information, including messages, reviews and support chats, to train machine learning models, and that it removes personally identifiable information from that text first. The company described the system in a post on its engineering blog published on 17 November, credited to four Airbnb engineers. [1]
The foundation is Presidio, an open-source library that Airbnb describes as originally developed by Microsoft and since moved to community governance. Airbnb says it picked Presidio over managed cloud services because it needed on-premises hosting, and over standalone named entity recognition models because those only cover part of the problem. The library runs in-process inside a shared ML workflow step, so no separate service holds the raw data. [1]
For names and street addresses, Airbnb says it replaced Presidio's default spaCy engine with recognizers built on GLiNER, a zero-shot model that accepts an entity description at inference time. The company says this improved precision for both entity types on its internal benchmark; GLiNER2 is being evaluated as a candidate. Airbnb-specific identifiers such as listing IDs and gift card numbers are caught with pattern recognizers, which the company extended to match multi-word context phrases like "gift card". [1]
Rather than leaving placeholder tokens, the system substitutes realistic fake values, which Airbnb argues keeps training data useful and avoids models learning the redaction pattern itself. Each fake is seeded from the real value plus a per-run secret, so the same input maps to the same fake within a run but differs across runs, and the company says it does not store a real-to-fake mapping. Structured identifiers are salted and hashed, coordinates truncated, and 64-bit internal IDs passed through format-preserving encryption. [1]
De-identified datasets are tagged, and Airbnb says the tag follows the data into its model registry, where it is designed to gate production deployment. The company says it prioritises recall over precision, reports metrics per entity type and language, and ships a new version only if no previously working language or entity type regresses. It did not publish any recall or precision figures. [1]
Airbnb says the work so far covers text only and that it has begun on de-identifying other media. [1]
Sources
- Airbnb: De-identifying text at scale: How Airbnb protects personal data in AI trainingSource published · TechDots found it
Statements labelled as a company's claim have not been independently confirmed by TechDots.
Get Technology stories by email
Related coverage
Spotify Engineering questions the idea of a single 'AI product manager' role
A Nov. 9 post on Spotify's engineering blog examines AI product management from three angles, starting from the observation that most coverage treats 'the AI PM' as one job. Only an excerpt was available for review.
Stripe says it moved $1B in vendor payments onto its own Global Payouts product
In a company blog post, Stripe says it began shifting its vendor payments to Global Payouts in early 2026, cutting new-market setup from six weeks to one with a 99.9% payout success rate. The figures are Stripe's own and have not been independently verified.
Stripe says users can now query their Stripe data in natural language through an agent
In a 5 November blog post, Stripe said customers can use natural language with an agent to analyse their Stripe data. Its examples include building an investor slide deck and a custom revenue dashboard. Details beyond the announcement are not available.