Ai4Privacy.
Menu

Detect and mask PII across your AI workflow

Find personal information in your text and replace it with placeholders. Ai4Privacy gives you the synthetic datasets, the browser-based chat, the REST API and the local SDKs to do it.

Explore open datasetsExplore products

See the prompt your AI should receive

Local preview
Sample prompt3 types detected

Original text: Draft a reply to Maya Chen at maya.chen@example.org about order EU-84921. Protected text: Draft a reply to [NAME_1] at [EMAIL_1] about order [ORDERNUMBER_1].

3 entities ready

Cited by

Ai4Privacy’s work is cited in research and industry publications from these institutions. Each logo links to its source.

Four ways to use Ai4Privacy

Training a model, working with prompts, shipping a product or running it locally. Compare all products, then choose the path that matches the work in front of you.

Choose a dataset release

Use synthetic, multilingual records for training and evaluation without introducing real personal data. Compare releases by scale, coverage, license and best fit.

PII Masking 2MEuropean Release

Scale
2M+ rows
Coverage
32 locales · 98 entities
License
CC-BY-4.0 + Enterprise
Best fit
European training and evaluation

PII Masking 1MSeries

Scale
1M series
Coverage
11 regions
License
CC-BY-4.0 open core
Best fit
Open-core model training

PII Masking 400Kand Below

Scale
400K+ largest release
Coverage
6 languages · 17–54 entities across the series
License
Ai4Privacy Dataset and Derivative Products License
Best fit
Comparing smaller release sizes

Inspect one verified OpenPII 1M row

Verified synthetic record · Hugging Face

View source on Hugging Face

Original synthetic record · exposed fields underlined

The set design, inspired by a miniature garden at Rectory Road 653 in Hungerford Kintbury, reflects the intimate scale of bonsai art.

language
en
region
GB
script
Latn
split
train
uid
20986105
One verified synthetic record from Ai4Privacy’s OpenPII 1M dataset. The interaction changes the view; it does not generate or modify the record.
Browse all datasets

Use PII masking directly

Use Chat, the REST API or local SDKs when you need masking capability directly rather than training and evaluation data.

Mask sensitive text before it reaches AI chat

Ai4Privacy Chat masks personal information in the browser before prompts are sent.

Ai4Privacy ChatSynthetic example
You

Draft a reply to [NAME_1] about order [ORDERNUMBER_1].

AI

Of course. What would you like the reply to say?

2 fields masked locally
Open Ai4Privacy Chat

Add PII detection and masking to your product

Call the REST API for detection, redaction, synthetic identities and privacy-aware chat.

  • Detection
  • Redaction
  • Synthetic identities
  • Privacy-aware chat
Request API access

Build a masking pipeline you can explain

Learn how to define entity types, prepare training data, evaluate models and apply masking decisions consistently.

PII Masking for AI Training Data: A Practitioner’s Guide

PII masking for AI training data is the process of finding personal data in a text corpus and replacing it, so that less identifiable information is carried into a model’s weights.

Read the practitioner’s guide
DefineInventoryChoose the personal-data types that matter.
FindDetectLocate each personal span in context.
ProtectReplaceKeep useful language around the placeholder.

Explore the complete guide set

PII Annotation Guidelines: Labeling Personal Data for NER Training

A PII annotation guideline is the written standard a labeler, or annotator, works from. It says which spans of text to mark, where each span starts and stops, and what to do when a case is ambiguous.

Read the annotation guide
False negative

Email maya@example.org today.

PII left exposed
False positive

Ship the [DATE_1] release.

Useful context removed
Two model-evaluation failure modes: a false negative leaves an email address exposed; a false positive masks the word May when it describes a software release.

How to Evaluate a PII Detection Model: Precision, Recall and What They Miss

Evaluating a personally identifiable information detection model means measuring how many personal data spans it finds, how many things it flags that are not personal data, and what the aggregate figure is hiding.

Read the evaluation guide

PII Entity Types: What a Detection Model Has to Recognize

A practical taxonomy of PII entity types, grouped by what they describe and how difficult they are for a detection model to recognize reliably.

Read the entity-types guide

How to Train a PII Detection Model with Named Entity Recognition

Build a PII detection model around explicit task framing, aligned labels, leak-free data splits and a clear decision about when training a new model is justified.

Read the training guide

Data Minimization for AI Training Data Under the GDPR

Map data-minimization principles to practical decisions about what a training corpus contains, what gets masked and what should not be collected at all.

Read the data-minimization guide

A shared language for personal data

p5y is Ai4Privacy’s open framework for describing personal and sensitive data. It is inspired by i18n and l10n.

English (en-US)

Ms. North, our office in Essex confirmed your visit with a health psychologist at 10:02 AM.

p5y redacted English (en-US-P)

[TITLE_1] [NAME_1], our office in [COUNTY_1] confirmed your visit with [MEDICAL_PROFESSIONAL_1] at [TIME_1].

Privacy masking treated as a translation task. Adapted from the p5y framework diagram.
  1. 01
    AwarenessDescribe what personal data is present.
  2. 02
    ProtectionApply a shared protection vocabulary.
  3. 03
    Quality assuranceCheck the result consistently.
Read the p5y specification

Adoption in practice

Research citations show relevance. These figures show how the work has been distributed and expanded by Artificial Intelligence Suisse SA and its partners.

3M+ · 30 · 3Examples, languages and regions

Ai4Privacy and VNCyberS expanded the Asia-Pacific release into a multilingual corpus spanning three regions.

Read the partner release 15M+downloads on Hugging FaceOpen the Ai4Privacy organization

Questions before you choose a path

Short answers to the practical questions teams ask when comparing datasets, local packages and hosted access.

What does Ai4Privacy do?

Ai4Privacy helps teams find personal information in text and replace it with placeholders. The same workflow is available through synthetic datasets, a browser-based chat, a REST API and local Python and JavaScript SDKs.

Do the training datasets contain real personal data?

The datasets presented here use synthetic PII rather than real personal data. Each release has its own coverage, access and license details, so those facts should be checked on the individual release page.

Can PII detection run inside my own application?

Yes. The Python and JavaScript packages are designed for local execution inside your own application or workflow. The REST API is the separate hosted integration path.

Start with our open-source work

The datasets are on Hugging Face, the projects are on GitHub, and the packages install from PyPI and npm.

If you are evaluating Ai4Privacy for your company, talk to us about access and integration.

Book an enterprise call